Check Your Website’s Ranking and Authority in DeepSeek AI Search.
Position
#1Rerank Score
2.41 / 3Authority
4 / 4Results for “GPT-6 Astra”
Relevance ≥ 1.5 · High quality
Authority ≥ 3 · Highly authoritative
All search results
| Rank | site_name | title | url | rerank_score / 3 | authority / 4 | snippet | summary | content | site_icon | dateLocal time |
|---|---|---|---|---|---|---|---|---|---|---|
| 1 | OpenAI Developers | GPT-6 Astra Model | OpenAI API | https://developers.openai.com/api/docs/models/gpt-6-astra | 2.41 | 4 | GPT-6 Astra is our most capable model for the most demanding work. Use it for complex reasoning, coding, computer use, research, and document creation.reasoning.effort supports low, medium, high, xhigh, and max. | GPT-6 Astra ... Our most capable model for the most demanding work. ... Reasoning ... Speed Fast Price $10•$50 Input•Output Input Text, Image ... GPT-6 Astra is our most capa…Show full textGPT-6 Astra ... Our most capable model for the most demanding work. ... Reasoning ... Speed Fast Price $10•$50 Input•Output Input Text, Image ... GPT-6 Astra is our most capable model for the most demanding work. Use it for complex reasoning, coding, computer use, research, and document creation.reasoning.effort supports low, medium, high, xhigh, and max. Get started with GPT-6 Astra using the model guide. 1,050,000 context window 128,000 max output tokens Apr 30, 2026 knowledge cutoff Reasoning token support ... Pricing is based on the number of tokens used, or other metrics based on the model type. For tool-specific models, like search and computer use, there’s a fee per tool call. See details in the pricing page. Text tokens Per 1M tokens Input $10.00 Cached input $1.00 Cache writes $12.50 ... $50.00 Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request. Cache writes are billed at 1.25x the uncached input token rate. Batch and Flex are priced at 50% of Standard rates. Fast mode is priced at 2x the applicable rates. Modalities Text Input and output Image Input only Audio Not supported Video ... Endpoints Live v1/live/sessions Chat Completions v1/chat/completions Responses v1/responses Realtime v1/realtime Realtime translation v1/realtime/translations Realtime transcription v1/realtime/transcription_sessions Assistants v1/assistants Batch v1/batch Fine-tuning v1/fine-tuning Embeddings v1/embeddings Image generation v1/images/generations ... v1/videos Image edit v1/images/edits Speech generation v1/audio/speech ... v1/audio/transcriptions ... v1/audio/translations Moderation v1/moderations Completions (legacy) ... Features Streaming ... Function calling ... Structured outputs ... Tools Tools supported by this model when using the Responses API. Web search ... File search ... Code interpreter ... Hosted shell ... Apply patch ... Skills ... MCP ... Tool search ... Snapshots Snapshots let you lock in a specific version of the model so that performance and behavior remain consistent. Below is a list of all available snapshots and aliases for GPT-6 Astra. gpt-6-astra ... Rate limits Rate limits ensure fair and reliable access to the API by placing specific caps on requests, tokens, audio duration, or other usage within a given time period. Your usage tier determines how high these limits are set and automatically increases as you send more requests and spend more on the API. Tier | RPM | TPM | Batch queue limit ... Free | Not supported Build | 5,000 | 1,000,000 | 3,000,000 Launch | 10,000 | 4,000,000 | 200,000,000 Grow | 15,000 | 40,000,000 | 15,000,000,000 | For the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to the page URL. GPT-6 Astra Default Our most capable…Show full textFor the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to the page URL. GPT-6 Astra Default Our most capable model for the most demanding work. Our most capable model for the most demanding work. Reasoning Highest Speed Fast Price $10•$50 Input•Output Input Text, Image Output Text GPT-6 Astra is our most capable model for the most demanding work. Use it for complex reasoning, coding, computer use, research, and document creation.reasoning.effort supports low, medium, high, xhigh, and max. Get started with GPT-6 Astra using the model guide. 1,050,000 context window 128,000 max output tokens Apr 30, 2026 knowledge cutoff Reasoning token support Pricing Pricing is based on the number of tokens used, or other metrics based on the model type. For tool-specific models, like search and computer use, there’s a fee per tool call. See details in the pricing page. Text tokens Per 1M tokens Input $10.00 Cached input $1.00 Cache writes $12.50 Output $50.00 Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request. Cache writes are billed at 1.25x the uncached input token rate. Batch and Flex are priced at 50% of Standard rates. Fast mode is priced at 2x the applicable rates. Modalities Text Input and output Image Input only Audio Not supported Video Not supported Endpoints Live v1/live/sessions Chat Completions v1/chat/completions Responses v1/responses Realtime v1/realtime Realtime translation v1/realtime/translations Realtime transcription v1/realtime/transcription_sessions Assistants v1/assistants Batch v1/batch Fine-tuning v1/fine-tuning Embeddings v1/embeddings Image generation v1/images/generations Videos v1/videos Image edit v1/images/edits Speech generation v1/audio/speech Transcription v1/audio/transcriptions Translation v1/audio/translations Moderation v1/moderations Completions (legacy) v1/completions Features Streaming Supported Function calling Supported Structured outputs Supported Fine-tuning Not supported Tools Tools supported by this model when using the Responses API. Web search Supported File search Supported Image generation Supported Code interpreter Supported Hosted shell Supported Apply patch Supported Skills Supported Computer use Supported MCP Supported Tool search Supported Snapshots Snapshots let you lock in a specific version of the model so that performance and behavior remain consistent. Below is a list of all available snapshots and aliases for GPT-6 Astra. gpt-6-astra gpt-6-astra gpt-6-astra Rate limits Rate limits ensure fair and reliable access to the API by placing specific caps on requests, tokens, audio duration, or other usage within a given time period. Your usage tier determines how high these limits are set and automatically increases as you send more requests and spend more on the API. Tier | RPM | TPM | Batch queue limit ---|---|---|--- Free | Not supported Build | 5,000 | 1,000,000 | 3,000,000 Launch | 10,000 | 4,000,000 | 200,000,000 Grow | 15,000 | 40,000,000 | 15,000,000,000 ## Docs agent Loading docs agent... | https://cdn.deepseek.com/site-icons/openai.com | Unknown |
| 2 | OpenAI Developers | Models | OpenAI API | https://developers.openai.com/api/docs/models | 2.35 | 4 | use GPT-6 Astra ... Start with GPT-6 Astra for complex reasoning and coding, choose GPT-6.1 Sol to balance intelligence and cost, or use GPT-6 Luna for cost-sensitive ... Our most …Show full textuse GPT-6 Astra ... Start with GPT-6 Astra for complex reasoning and coding, choose GPT-6.1 Sol to balance intelligence and cost, or use GPT-6 Luna for cost-sensitive ... Our most capable model for the most demanding work. Model ID gpt-6-astra ... $50 / Output MTok Max output 128K tokens Context window 1.05M | If you're not sure where to start, use GPT-6 Astra, our flagship model for complex reasoning and coding. ... Start with GPT-6 Astra for complex reasoning and coding, choose GPT-6.1…Show full textIf you're not sure where to start, use GPT-6 Astra, our flagship model for complex reasoning and coding. ... Start with GPT-6 Astra for complex reasoning and coding, choose GPT-6.1 Sol to balance intelligence and cost, or use GPT-6 Luna for cost-sensitive, high-volume workloads. ... Our most capable model for the most demanding work. Model ID gpt-6-astra lowmediumhighxhighmax Input price $10 / Input MTok Output price $50 / Output MTok Max output 128K tokens Context window 1.05M Knowledge cutoff Apr 30, 2026 | For the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to the page URL. # Models ## Choosing a model If you'…Show full textFor the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to the page URL. # Models ## Choosing a model If you're not sure where to start, use GPT-6 Astra, our flagship model for complex reasoning and coding. Choose GPT-6.1 Sol to balance intelligence and cost, or GPT-6 Luna for cost-sensitive, high-volume workloads. All latest OpenAI models support text and image input, text output, multilingual capabilities, and vision. Models are available via the Responses API and our Client SDKs. ## Flagship models Start with GPT-6 Astra for complex reasoning and coding, choose GPT-6.1 Sol to balance intelligence and cost, or use GPT-6 Luna for cost-sensitive, high-volume workloads. GPT-6 Astra Our most capable model for the most demanding work. Model ID gpt-6-astra lowmediumhighxhighmax Input price $10 / Input MTok Output price $50 / Output MTok Max output 128K tokens Context window 1.05M Knowledge cutoff Apr 30, 2026 GPT-6.1 Sol Near-Astra performance for complex work at a lower cost. Model ID gpt-6.1-sol lowmediumhighxhighmax Input price $2 / Input MTok Output price $10 / Output MTok Max output 128K tokens Context window 1.05M Knowledge cutoff Apr 30, 2026 GPT-6 Luna Our most efficient model for focused, high-volume tasks. Model ID gpt-6-luna nonelowmediumhighxhighmax Input price $0.1 / Input MTok Output price $0.5 / Output MTok Max output 128K tokens Context window 1.05M Knowledge cutoff May 18, 2026 ## Specialized models Purpose-built for specific tasks. Life sciences Models for life sciences research Speech generation Models for generating natural-sounding speech from text ## Browse our full catalog of models Diverse models for a variety of tasks ## Docs agent Loading docs agent... | https://cdn.deepseek.com/site-icons/openai.com | Unknown |
| 3 | OpenAI Developers | Using GPT-6 | OpenAI API | https://developers.openai.com/api/docs/guides/latest-model | 2.32 | 4 | GPT-6 Astra tends to use lists, tables and Markdown to make responses scannable. ... GPT-6 Astra is trained to be able to divide and delegate work to subagents that work in paralle…Show full textGPT-6 Astra tends to use lists, tables and Markdown to make responses scannable. ... GPT-6 Astra is trained to be able to divide and delegate work to subagents that work in parallel. ... GPT-6 Astra ... GPT-6 Astra ... Chat Completions, but tool calling requires Responses. | Use this prompt to find silent and conflicting guidance when your application loads many skills and instruction files such asAGENTS.md. ### Personality and writing style GPT-6 As…Show full textUse this prompt to find silent and conflicting guidance when your application loads many skills and instruction files such asAGENTS.md. ### Personality and writing style GPT-6 Astra tends to use lists, tables and Markdown to make responses scannable. If your application needs prose with less formatting, specify that preference. Default to using clear, concise paragraphs, each developing one main idea. Use lists only when the information is genuinely parallel, sequential, or easier to compare, and avoid nested lists unless the hierarchy cannot be expressed clearly in prose. Use plain, simple language: familiar words, concrete examples, and precise verbs. Prefer active voice and direct statements. Make sure to state the main point clearly and early, then develop it with the explanation and detail the reader needs. Let each sentence build on what came before. Develop the points that matter and provide enough support to be useful. For technical communication, the following prompt helps strike a balance between using clear, coherent language while remaining domain appropriate: Use plain language over jargon, and reference technical details only to the degree that it helps illustrate an idea or your work to the user. Communicate complex concepts in a clear and cohesive manner, and calibrate your writing to the level of background knowledge assumed from the user's prompt and context. To reduce jargon and stock phrases in writing, start with this prompt: Avoid using slop words or phrases like "Bottom Line:" in conclusions, "delve," "foster," "leverage," "it's worth noting," "importantly," "Question? Answer." or "This isn't about X. It's about Y.", "genuinely" or hyphenated compound descriptions and adjectives. Do not use concluding summary statements such as "In short:..", "The simplest mental model is:...". State the intended action directly. Avoid adding what you won't do, what will remain unchanged, or how you'll separate or categorize results. Do not use contrastive framing such as "X, not Y" that introduces an unprompted alternative that the user didn't ask about. Avoid invented compound labels like "exact-head checks" and "editorial-row layouts", vague qualifiers, and canned transitions; use plain verbs and prepositions to state the actual relationship directly. ### Subagent delegation GPT-6 Astra is trained to be able to divide and delegate work to subagents that work in parallel. If you are implementing a multi-agent system in your harness, use the following prompt to tune how much GPT-6 Astra should delegate work: If at any point you can parallelize work by delegating tasks to another agent (no matter if you are the root or subagent), you should do so using collaboration tools if it could save time or improve quality. Messages between agents may contain grammar or spacing errors. Use this prompt to make inter-agent messages easier to read: Messages that you send to other agents and your final answer may be read by a human, so ensure they are legible. Always put proper spaces between words and/or numbers. The model tends to respond well to prompting for how and when it should delegate work to subagents, so tune this behavior to fit with your harness and multi-agent implementation. ### Testing and verification For coding tasks, calibrate how much testing and verification a change requires. This can help avoid unnecessary tests or repeated checks for small changes. Do not write tests for reversible, low-impact changes that mirror the implementation. If you do choose to verify your work with tests, make sure that the tests are meaningful and necessary to verify implementation. Run tests appropriate to the change and complete required checks. Once those pass, broaden or repeat testing only when new changes, failures, or unresolved concerns justify it; otherwise, continue toward completing the task. ## Migration quickstart ### Migrate with Codex Codex can apply the recommended changes in this guide with the OpenAI Docs skill. $openai-docs migrate this project to the GPT-6 model family To use this skill in other coding agents, download it from the Codex repository. ### Update API and model parameters Setmodel to gpt-6-astra, gpt-6.1-sol, or gpt-6-luna, then check the following: - Reasoning effort: Preserve your current effective reasoning effort where supported. GPT-6 Astra and GPT-6.1 Sol do not supportnone; use low instead. GPT-6 Sol and GPT-6 Luna support none. If your existing request uses minimal, start with low and compare results on representative tasks. Use reasoning.effort in Responses or reasoning_effort in Chat Completions. - Tool calling: Use the Responses API. GPT-6 Astra and GPT-6.1 Sol support Chat Completions, but tool calling requires Responses. GPT-6 Sol... | For the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to the page URL. # Using GPT-6 Learn best practices, f…Show full textFor the complete documentation index, see llms.txt. Markdown versions of documentation pages are available by appending .md to the page URL. # Using GPT-6 Learn best practices, features, and migration guidance for GPT-6 Astra and the GPT-6 model family. Choose a GPT-6 model based on the reasoning your task requires, speed, and cost. Highest intelligence For the most demanding reasoning, coding, and professional work. Balanced speed, cost, and intelligence Near-Astra performance for complex work at a lower cost. Fastest and most cost-effective Strong performance for focused, high-volume tasks. To get started, setmodel in a Responses API request. If you already use gpt-6-sol, review the migration guidance before switching to GPT-6.1 Sol. ### GPT-6 Astra GPT-6 Astra is our most intelligent model yet, with state-of-the-art performance in computer use, browsing, software engineering, science, and professional work. It can carry out multi-step workflows across code, browsers, and professional software. In several evaluations, Astra achieved stronger results using substantially fewer output tokens. Its estimated API cost per task was lower than earlier models despite its higher per-token pricing. GPT-6 Astra is also our most aligned model yet. It excels at exercising care, respecting task boundaries, and communicating transparently. When instructions leave room for interpretation, it uses the context it has to fill in routine gaps and asks focused questions when the answer could change the outcome. It incorporates new requirements, changes course when asked, and answers side questions without losing track of the broader task. All GPT-6 Astra users also have access to Fast mode and the new Ultrafast mode for our fastest API speeds. ### GPT-6.1 Sol Use GPT-6.1 Sol for complex coding, computer use, and professional work when you want near-Astra performance at a lower cost. Compare it with Astra on your tasks to assess the tradeoff between quality and cost. Setreasoning.effort to low, medium (default), high, xhigh, or max. Use the Responses API for tool calling. Chat Completions supports requests without tools. The none and minimal reasoning efforts are not supported. See the model page for specifications, pricing, and availability, or model selection for guidance on choosing a model. ## What’s new - Async tool calling: GPT-6 can continue reasoning, call other tools, or answer independent parts of a request while your application runs a tool. Setasync: true on a function or custom tool and return its result when ready using the original call_id. Your application still executes the tool and manages pending work. See Async tool calling for basic usage and a developer-defined wait-tool pattern. - Mid-turn steering: Send additional user instructions while GPT-6 is working, such as a correction or a change in requirements. Over a WebSocket connection, the Responses API preserves completed work and includes the update in a continuation. See Mid-turn steering for the event flow and tool-result handling. - Change reasoning mid-conversation while preserving cache: Add aconfiguration_update input item to increase reasoning effort for difficult work or reduce it for routine follow-ups without rewriting the original prompt prefix. The updated reasoning effort applies until another configuration_update input item overrides it. See Change reasoning mid-conversation for examples and compatibility. - Misalignment monitoring: As part of our strengthened safeguards for GPT-6 Astra, our systems asynchronously monitor for misalignment and trigger alerts when necessary. See Misalignment monitoring for more information. GPT-6 also supports the existing API capabilities available with GPT-5.6, including computer use, Structured Outputs, streaming, Programmatic Tool Calling, multi-agent orchestration, prompt caching, persisted reasoning, compaction, and pro mode. ## Limitations - GPT-6 Astra and GPT-6.1 Sol do not support thenone reasoning effort; GPT-6 Sol and GPT-6 Luna do. - Fast mode supports EU data residency for GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna, but not GPT-6 Astra. Ultrafast mode for GPT-6.1 Sol supports US and EU data residency and global processing. GPT-6 Astra Ultrafast supports US data residency and global processing only. See data residency eligibility. ## Prompting best practices Use the following prompts as a starting point across the GPT-6 model family. They address behavior observed with GPT-6 Astra; evaluate them with your chosen model and workload. ### GPT-6 Astra behavior - Initiative and follow-through: The model is designed to be a more effective collaborator and is thus more likely to ask the user a question when additional input could materially change the result. This can cause it to stop when the user may expect it to make reasonable assumptions and persist. - Instruction following: GPT-6 Astra is stronger at general instruction following than our previous models, giving you greater control over its behavior. It can be more sensitive to instructions contained in skills and other files, such asAGENTS.md. We strongly recommend auditing skills and other files accessible to your model for instructions that could influence its behavior. - Personality and writing style: The model tends toward detailed, formatted responses and may use recurring phrases across sessions. Specify the writing style and structure your application needs. - Subagent delegation: The model may delegate less often than desired for your workflow. Specify when and how much it should use subagents for parallel work. - Testing and verification: For coding tasks, the model tends to be thorough in testing before considering a task complete. For smaller tasks, this can result in broader tests than the task requires. ### Initiative and follow-through GPT-6 Astra is generally better than GPT-5.6 Sol and earlier models at staying coherent during long tasks. It is also more likely to ask for clarification where earlier models would make assumptions. To encourage more autonomous work, start with this prompt: You should infer the user's intent and task scope from the instructions and prior conversation context. Your job is to bias towards action and carry the user's intended task to completion. When the user expresses intent to perform new work or fix an existing issue, persist until the user's intended goal is complete. Progress autonomously towards the user's goal (e.g. creating isolated worktrees / checkouts if needed, resolving merge conflicts, read-only actions, creating draft PRs etc.) unless they are clearly destructive or irreversible. When the user’s intent is unclear, the model is more likely to ask the user for clarification to proceed. Prompt the model to follow through if the user’s prompt implies authorization: When the user's prompt indicates a request for action, such as "can you...", "I want to...", "help me..." and similar expressions, treat these as instructions to do the work and take action. Do not stop at acknowledging capability (e.g. "Yes…"), proposing a plan, or offering to continue. Do not settle for a partial or "helpful enough" solution that does not fully satisfy the user's task to save time, effort or tokens. If a task requires sustained work, complete all the necessary work until the intended outcome is fulfilled. Prompt the model to ask for approval only after preparing a concrete, reviewable result. This avoids blocking the task before the model has done the work it can, and often leads to quicker task completion. Before asking the user clarifying questions, you should complete the work that is already authorized from context and necessary to make the proposed action concrete and reviewable. The user should be approving a concrete, reviewable result. For example, before deploying a change, writing to an external application, merging a PR or publishing a site, do all the required work first so that user approval is the final step. You don't need user permission for reversible tasks, read-only actions, reviews or fixes, or anything for which authorization is provided earlier in the session or strongly implied from the task instruction. Do not introduce unsolicited warnings, disclaimers, approval flows, or safety/compliance checklists due to hypothetical risk. The model also likes to ask non-blocking questions as it’s working by default, so adjust these prompts to match the level of autonomy your application needs. ### Instruction following GPT-6 Astra is better able to follow longer instructions, but can also be more sensitive to information in context. For example, unclear or conflicting guidance in a skill file may cause the model to pause and block work early. Make the priority of user instructions and skills explicit. The user's instructions take precedence over guidelines provided in a skill. If explicit user instructions conflict with a skill's instructions, prioritize the user's instructions. Asking the model to identify the skill and instruction that caused it to pause or change direction can also be effective in providing transparency into model behavior. If a skill causes you to ask for permission or confirmation, pause, leave requested work unfinished, or diverge from the user's intent, name and link to the exact SKILL.md file you read, quote the relevant instruction, and briefly explain how it applies. Distinguish explicit skill requirements from your interpretation of guidelines. Use this prompt to find silent and conflicting guidance when your application loads many skills and instruction files such asAGENTS.md. ### Personality and writing style GPT-6 Astra tends to use lists, tables and Markdown to make responses scannable. If your application needs prose with less formatting, specify that preference. Default to using clear, concise paragraphs, each developing one main idea. Use lists only when the information is genuinely parallel, sequential, or easier to compare, and avoid nested lists unless the hierarchy cannot be expressed clearly in prose. Use plain, simple language: familiar words, concrete examples, and precise verbs. Prefer active voice and direct statements. Make sure to state the main point clearly and early, then develop it with the explanation and detail the reader needs. Let each sentence build on what came before. Develop the points that matter and provide enough support to be useful. For technical communication, the following prompt helps strike a balance between using clear, coherent language while remaining domain appropriate: Use plain language over jargon, and reference technical details only to the degree that it helps illustrate an idea or your work to the user. Communicate complex concepts in a clear and cohesive manner, and calibrate your writing to the level of background knowledge assumed from the user's prompt and context. To reduce jargon and stock phrases in writing, start with this prompt: Avoid using slop words or phrases like "Bottom Line:" in conclusions, "delve," "foster," "leverage," "it's worth noting," "importantly," "Question? Answer." or "This isn't about X. It's about Y.", "genuinely" or hyphenated compound descriptions and adjectives. Do not use concluding summary statements such as "In short:..", "The simplest mental model is:...". State the intended action directly. Avoid adding what you won't do, what will remain unchanged, or how you'll separate or categorize results. Do not use contrastive framing such as "X, not Y" that introduces an unprompted alternative that the user didn't ask about. Avoid invented compound labels like "exact-head checks" and "editorial-row layouts", vague qualifiers, and canned transitions; use plain verbs and prepositions to state the actual relationship directly. ### Subagent delegation GPT-6 Astra is trained to be able to divide and delegate work to subagents that work in parallel. If you are implementing a multi-agent system in your harness, use the following prompt to tune how much GPT-6 Astra should delegate work: If at any point you can parallelize work by delegating tasks to another agent (no matter if you are the root or subagent), you should do so using collaboration tools if it could save time or improve quality. Messages between agents may contain grammar or spacing errors. Use this prompt to make inter-agent messages easier to read: Messages that you send to other agents and your final answer may be read by a human, so ensure they are legible. Always put proper spaces between words and/or numbers. The model tends to respond well to prompting for how and when it should delegate work to subagents, so tune this behavior to fit with your harness and multi-agent implementation. ### Testing and verification For coding tasks, calibrate how much testing and verification a change requires. This can help avoid unnecessary tests or repeated checks for small changes. Do not write tests for reversible, low-impact changes that mirror the implementation. If you do choose to verify your work with tests, make sure that the tests are meaningful and necessary to verify implementation. Run tests appropriate to the change and complete required checks. Once those pass, broaden or repeat testing only when new changes, failures, or unresolved concerns justify it; otherwise, continue toward completing the task. ## Migration quickstart ### Migrate with Codex Codex can apply the recommended changes in this guide with the OpenAI Docs skill. $openai-docs migrate this project to the GPT-6 model family To use this skill in other coding agents, download it from the Codex repository. ### Update API and model parameters Setmodel to gpt-6-astra, gpt-6.1-sol, or gpt-6-luna, then check the following: - Reasoning effort: Preserve your current effective reasoning effort where supported. GPT-6 Astra and GPT-6.1 Sol do not supportnone; use low instead. GPT-6 Sol and GPT-6 Luna support none. If your existing request uses minimal, start with low and compare results on representative tasks. Use reasoning.effort in Responses or reasoning_effort in Chat Completions. - Tool calling: Use the Responses API. GPT-6 Astra and GPT-6.1 Sol support Chat Completions, but tool calling requires Responses. GPT-6 Sol and GPT-6 Luna support function calling in Chat Completions only withreasoning_effort: "none". Use Responses for reasoning with tools. - Unsupported parameters: When reasoning effort is notnone, remove temperature, top_p, and top_logprobs. For Chat Completions, also remove logprobs. For Responses, remove message.output_text.logprobs from include. - Data residency: Fast mode supports EU data residency for GPT-6.1 Sol, GPT-6 Sol, and GPT-6 Luna, but not GPT-6 Astra. Ultrafast mode for GPT-6.1 Sol supports US and EU data residency and global processing. GPT-6 Astra Ultrafast supports US data residency and global processing only. Fast mode for GPT-6 Astra does not include a latency SLA. See Fast mode compatibility. - Changing reasoning effort: If your application changes effort between responses, useconfiguration_update items in standard, single-agent requests. Keep request-level reasoning.effort unchanged to preserve the prompt prefix for caching. Check the compatibility limits before adopting this feature. - Prompt caching: When migrating from GPT-5.5 or earlier, replaceprompt_cache_retention with prompt_cache_options.ttl set to "30m". Review the prompt caching changes, including cache boundaries and cache-write billing. - Unnecessary approval pauses: If you run into issues where the model keeps asking for approval before proceeding, use the initiative and follow-through guidance to prompt for more autonomous execution. See the rest of Prompting best practices for guidance on instruction following, writing style, subagent delegation, and testing. ## Previous models ## Using GPT-5.6 Learn best practices, features, and migration guidance for GPT-5.6 and the GPT-5.6 model family. ## Introduction GPT-5.6 sets a new quality and efficiency baseline for complex production workflows. GPT-5.6 is especially token-efficient and improves frontend aesthetics, including layout, visual hierarchy, and design judgment. GPT-5.6 also introduces a new naming scheme. Thegpt-5.6 alias routes requests to gpt-5.6-sol, the model for flagship capability. Use gpt-5.6-terra for strong performance at a lower price and gpt-5.6-luna for efficient, high-volume workloads. When migrating from GPT-5.5 or GPT-5.4, start with your current GPT-5.5 or GPT-5.4 reasoning setting, then test the same setting and one level lower on representative tasks. GPT-5.6 can often maintain or improve quality with fewer tokens, but the best setting depends on your workload. ## What’s new - Programmatic Tool Calling: GPT-5.6 can write JavaScript to call eligible tools, pass results between calls, and process intermediate outputs in a hosted runtime. Use Programmatic Tool Calling for bounded, tool-heavy workflows that do not require fresh model judgment between each step. Programmatic Tool Calling is ZDR-compatible with no additional container costs. - Multi-agent [beta]: Multi-agent lets a GPT-5.6 instance coordinate multiple subagents in parallel and synthesize their results. Similar to ultra mode in Codex, this can reduce wall-clock time and improve performance for complex tasks that divide cleanly into independent workstreams. Multi-agent is available as a beta feature in the Responses API as we iterate on developer feedback. - Explicit prompt caching: GPT-5.6 lets you mark exactly which reusable prompt prefixes OpenAI caches. You can still use automatic caching in implicit mode. OpenAI bills cache writes at 1.25× the uncached input rate, while cache reads remain discounted. Learn how to configure prompt caching. - Persisted reasoning: GPT-5.6 can reuse available reasoning items across turns to improve multi-turn quality and cache efficiency. Usereasoning.context to select the behavior. Learn how to preserve reasoning across calls. - Max reasoning effort: GPT-5.6 supportsmax reasoning effort for demanding tasks that need more exploration and verification. If you currently use xhigh, compare both settings on representative workloads. - Pro mode: GPT-5.6 can perform more model work to improve reliability on difficult tasks and return a single final answer. Enable it withreasoning.mode: "pro" when quality matters more than latency and token usage. Learn how to use pro mode. - Token efficiency: GPT-5.6 reaches flagship-level performance with fewer output tokens. - Frontend design: GPT-5.6 creates more polished and usable websites and applications, with stronger layout, visual hierarchy, and design judgment. - Intent understanding: GPT-5.6 can better infer the user’s underlying goal and intended level of work from context, so you often do not need to prescribe every step. Continue to provide domain context, hard constraints, approval boundaries, and success criteria. Tell the model when an important ambiguity should trigger a question. - Original image detail: GPT-5.6 preserves image dimensions withoriginal or auto detail, except that images larger than 65,535 pixels on either side are scaled down to fit that limit. The API rejects images that still exceed the 30,000-patch limit, rather than resizing them to fit it. Large images can use more input tokens and increase latency. Learn how to choose an image detail level. ## Safeguards When using GPT-5.6 models, users may encounter safeguards that block or refuse some requests due to real-time cyber and biology misuse classifiers that are run as model outputs are generated. Other requests may take longer because generation is paused for several seconds mid-stream while these classifiers synchronously review outputs. Safeguards may occasionally intervene on legitimate work, particularly in dual-use areas where defensive and offensive activity can initially look similar. If your application serves individual end users, send a stable, privacy-preservingsafety_identifier with each request. See Implement safety identifiers for guidance. We are continuously evolving these safeguards so that they are robust and effective in holding up to adversarial pressure, while preserving access to legitimate work such as code review, vulnerability research, patch development, debugging, security education, and defensive testing. ## Migration quickstart ### Migrate with Codex Codex can apply the recommended | https://cdn.deepseek.com/site-icons/openai.com | Unknown |
| 4 | 디지털투데이 | OpenAI shares case of automating software maintenance using GPT-6 | https://www.digitaltoday.co.kr/en/view/115997/openai-shares-case-of-automating-software-maintenance-using-gpt-6 | 2.29 | 3 | The company assigned GPT-6 Astra to identify and remove unused code by reviewing pull requests, Slack conversations and internal records. Engineers merged more than 1 ... OpenAI te…Show full textThe company assigned GPT-6 Astra to identify and remove unused code by reviewing pull requests, Slack conversations and internal records. Engineers merged more than 1 ... OpenAI teams then gave GPT-6 Astra short and vague tasks, asking it to find needed materials on Slack or the experiment platform Statsig and clean up unused code. | # OpenAI shares case of automating software maintenance using GPT-6 ## AI Summary OpenAI has disclosed on its blog how it automated internal software maintenance using its coding …Show full text# OpenAI shares case of automating software maintenance using GPT-6 ## AI Summary OpenAI has disclosed on its blog how it automated internal software maintenance using its coding agent Codex and GPT-6 family models. The company assigned GPT-6 Astra to identify and remove unused code by reviewing pull requests, Slack conversations and internal records. Engineers merged more than 1,400 pull requests and deleted over 800,000 lines of code. OpenAI also used Astra to clean up old experiments and feature flags, and tested cheaper GPT-6.1 Sol for similar tasks. - Writer Chi-gyu Hwang - published 2026-10-18 21:08:00 - updated 2026-10-18 21:08:03 OpenAI has disclosed on its blog a case in which it automated internal software maintenance using the coding agent Codex and GPT-6 family models. The company said development sped up thanks to the coding agent, increasing the maintenance burden. It said there was more work deleting unused code and cleaning up and fixing completed tests. OpenAI teams then gave GPT-6 Astra short and vague tasks, asking it to find needed materials on Slack or the experiment platform Statsig and clean up unused code. Astra reviewed past pull requests, Slack conversations, internal project records and performance data from production environments to select code to delete. It also explained the reasons for deletion. Five engineers incorporated more than 1,400 pull requests over two weeks while handling other tasks. More than 800,000 lines of code were deleted. One service cut its container image size by 48 percent to 1.72 GB from 3.31 GB. OpenAI also deployed Astra to clean up more demanding old experiments. Astra assessed CPU usage by experiment and found sources of waste that no one had known about for years. The team deleted hundreds of experiments and feature flags. Astra confirmed that one flag marked as "permanent" used a lot of CPU and requested human review. After the review, the team deleted the flag and cut CPU usage on major services by about 0.42 percent. The team also tested the same work with the cheaper GPT-6.1 Sol. After checking 6,000 files and 1 million lines of code, Sol set to the highest reasoning intensity found 131 issues, or 92 percent of the 142 issues Astra found. API costs were $15.12, or 14 percent of Astra's $109.06. The company also highlighted improvements in continuous integration, or CI. After incorporating about 30 pull requests based on Codex suggestions, the median CI completion time fell to about 4.5 minutes from about 10 minutes. Runtime for a bundle of 63 tests dropped 95 percent to 29.8 seconds from 592.6 seconds. The company advised starting with small maintenance tasks, reviewing results and then expanding the scope. | # OpenAI shares case of automating software maintenance using GPT-6 ## AI Summary OpenAI has disclosed on its blog how it automated internal software maintenance using its coding …Show full text# OpenAI shares case of automating software maintenance using GPT-6 ## AI Summary OpenAI has disclosed on its blog how it automated internal software maintenance using its coding agent Codex and GPT-6 family models. The company assigned GPT-6 Astra to identify and remove unused code by reviewing pull requests, Slack conversations and internal records. Engineers merged more than 1,400 pull requests and deleted over 800,000 lines of code. OpenAI also used Astra to clean up old experiments and feature flags, and tested cheaper GPT-6.1 Sol for similar tasks. - Writer Chi-gyu Hwang - published 2026-10-18 21:08:00 - updated 2026-10-18 21:08:03 OpenAI has disclosed on its blog a case in which it automated internal software maintenance using the coding agent Codex and GPT-6 family models. The company said development sped up thanks to the coding agent, increasing the maintenance burden. It said there was more work deleting unused code and cleaning up and fixing completed tests. OpenAI teams then gave GPT-6 Astra short and vague tasks, asking it to find needed materials on Slack or the experiment platform Statsig and clean up unused code. Astra reviewed past pull requests, Slack conversations, internal project records and performance data from production environments to select code to delete. It also explained the reasons for deletion. Five engineers incorporated more than 1,400 pull requests over two weeks while handling other tasks. More than 800,000 lines of code were deleted. One service cut its container image size by 48 percent to 1.72 GB from 3.31 GB. OpenAI also deployed Astra to clean up more demanding old experiments. Astra assessed CPU usage by experiment and found sources of waste that no one had known about for years. The team deleted hundreds of experiments and feature flags. Astra confirmed that one flag marked as "permanent" used a lot of CPU and requested human review. After the review, the team deleted the flag and cut CPU usage on major services by about 0.42 percent. The team also tested the same work with the cheaper GPT-6.1 Sol. After checking 6,000 files and 1 million lines of code, Sol set to the highest reasoning intensity found 131 issues, or 92 percent of the 142 issues Astra found. API costs were $15.12, or 14 percent of Astra's $109.06. The company also highlighted improvements in continuous integration, or CI. After incorporating about 30 pull requests based on Codex suggestions, the median CI completion time fell to about 4.5 minutes from about 10 minutes. Runtime for a bundle of 63 tests dropped 95 percent to 29.8 seconds from 592.6 seconds. The company advised starting with small maintenance tasks, reviewing results and then expanding the scope. About the Author Chi-gyu Hwang delight@d-today.co.kr | https://cdn.deepseek.com/site-icons/digitaltoday.co.kr | |
| 5 | 36Kr | ChatGPT推出AI小队功能 GPT-6 Astra Pro重磅升级 支持分工协作+进度查询 | https://eu.36kr.com/zh/p/4031321049715076 | 2.24 | 3 | 按照官方帮助文档，在模型菜单选择GPT-6，再把Thinking滑块拉到Pro档，用的就是由GPT-6 Astra驱动的GPT-6 Pro ... GPT-6的普通档位由Sol提供支持，Pro档由Astra提供支持。 | # ChatGPT有了AI小队，GPT-6 Astra Pro升级，分头干活还能查进度 新智元2026-10-18 17:45 ... 10月16日，OpenAI宣布升级ChatGPT中的GPT-6 Astra Pro： 回复更快，插件和工具使用能力增强，还能让多个智能体分头处理同一项任务。 一项复杂工作交出去，几个AI可以各管一块，同时推进。 你还…Show full text# ChatGPT有了AI小队，GPT-6 Astra Pro升级，分头干活还能查进度 新智元2026-10-18 17:45 ... 10月16日，OpenAI宣布升级ChatGPT中的GPT-6 Astra Pro： 回复更快，插件和工具使用能力增强，还能让多个智能体分头处理同一项任务。 一项复杂工作交出去，几个AI可以各管一块，同时推进。 你还能在对话里跟踪各自进度，单独点开已完成的任务，检查成果。 分头执行、跟踪进度、查看成果，这几个环节被接进了同一个聊天入口。 目前，这次更新正陆续向符合条件的Pro用户开放。公司或团队账号能否用上，还要看管理员设置，以及是否已获得推送。 ... ## 这支AI小队怎么开工 入口在ChatGPT的Chat标签页。 按照官方帮助文档，在模型菜单选择GPT-6，再把Thinking滑块拉到Pro档，用的就是由GPT-6 Astra驱动的GPT-6 Pro。 Instant、Medium、High和Extra High这几个档位，则由GPT-6 Sol驱动。 GPT-6的普通档位由Sol提供支持，Pro档由Astra提供支持。 这次升级，主要改了三件事。 首先是提速。 OpenAI表示，回复更快了。至于具体快多少、完成一项任务能省多久，官方暂未公布具体数字。 其次是工具使用。 AI能不能把活儿做完，也取决于工具用得顺不顺。 ... 这次OpenAI宣布改进插件和工具的使用能力，效果如何还要看「疗效」。 如果工具需要访问外部服务，用户仍要先连接账号、完成授权，AI只能在获准的权限范围内操作。 最核心的变化，是多个智能体并行干活。 比如，你想从三款AI工具里挑一款，需要比较功能、价格和使用限制。 AI可以分头搜集各款产品的资料，再把月付、年付、免费版和付费版的条件对齐。 ... 不过，AI分工和人类团队一样，能同时做的分头做，前后有依赖的则要按顺序来。 比如，先读材料、再定方案、最后修改文件，后一步要用到前一步的结果，多派几个AI也不能跳过。 ... 活儿分出去，进度也要能看得见。 用户可以在主对话里跟踪各个智能体，完成后再点开检查成果。 ... 至于最多能同时跑几个智能体、分工是否默认触发、结果具体怎样合并，这条公告中还没有展开说明。 ... ## 但也要看清额度 能不能用上、能用多少，是用户接下来要算的账。 根据官方帮助说明，Plus用户可以在Work和Codex中使用Astra，但不能在Chat里使用GPT-6 Pro。 这次Chat中的多智能体升级，目前正向符合条件的Pro用户逐步推送，Plus用户暂时还用不上。 ... Chat里的Pro模型消息限额，与Work、Codex共享的用量池分开计算。 上面这张用户截图也写明，共享用量不包含Chat对话。 然后，再看订阅哪一档。 当前官方帮助文档显示，200美元档Pro的GPT-6 Pro有周限额；100美元档Pro中，GPT-6 Pro和GPT-5.6 Sol Pro共用一份周额度。 ... 100美元档两款Pro模型共用周额度。 ... 一支AI小队好不好用，要看从接到任务到交出合格成果，总共花了多少时间、消耗了多少额度，以及还需要人做多少检查和修改。 ... https://x.com/OpenAI/status/2111192259073094035 https://help.openai.com/en/articles/6825453-chatgpt-release-notes https: ... https: ... 本文来自微信公众号“新智元”（ID：AI_era），作者：ASI启示录，编辑：元宇,36氪经授权发布。 | # ChatGPT有了AI小队，GPT-6 Astra Pro升级，分头干活还能查进度 新智元2026-10-18 17:45 ChatGPT有了AI小队！GPT-6 Astra Pro升级，分头干活还能查进度 ChatGPT也学会「摇人」组队干活了！ 10月16日，OpenAI宣布升级ChatGPT中的GPT-6 Astra Pro： 回复更快，…Show full text# ChatGPT有了AI小队，GPT-6 Astra Pro升级，分头干活还能查进度 新智元2026-10-18 17:45 ChatGPT有了AI小队！GPT-6 Astra Pro升级，分头干活还能查进度 ChatGPT也学会「摇人」组队干活了！ 10月16日，OpenAI宣布升级ChatGPT中的GPT-6 Astra Pro： 回复更快，插件和工具使用能力增强，还能让多个智能体分头处理同一项任务。 一项复杂工作交出去，几个AI可以各管一块，同时推进。 你还能在对话里跟踪各自进度，单独点开已完成的任务，检查成果。 分头执行、跟踪进度、查看成果，这几个环节被接进了同一个聊天入口。 目前，这次更新正陆续向符合条件的Pro用户开放。公司或团队账号能否用上，还要看管理员设置，以及是否已获得推送。 替你干活的AI多了，需要你盯的任务也多了。 网友@Sebasti54919704就吐槽：并行智能体很好，但挨个跟踪进度，才是让他崩溃的部分。 他索性给自己的Claude Code和Codex会话接上语音，开口就能询问进展、切换会话。 有时候，这些智能体还会反驳他。 几个AI同时开工，人却可能忙着盯进度、催任务，偶尔还得应付AI的「顶嘴」。 AI小队好不好用，也要看它能不能让用户少花时间追问进度、传递信息、协调任务。 ## 这支AI小队怎么开工 入口在ChatGPT的Chat标签页。 按照官方帮助文档，在模型菜单选择GPT-6，再把Thinking滑块拉到Pro档，用的就是由GPT-6 Astra驱动的GPT-6 Pro。 Instant、Medium、High和Extra High这几个档位，则由GPT-6 Sol驱动。 GPT-6的普通档位由Sol提供支持，Pro档由Astra提供支持。 这次升级，主要改了三件事。 首先是提速。 OpenAI表示，回复更快了。至于具体快多少、完成一项任务能省多久，官方暂未公布具体数字。 其次是工具使用。 AI能不能把活儿做完，也取决于工具用得顺不顺。 网友@Rossisonx就吐槽，插件有时会陷入循环，卡在原地，似乎还在消耗额度。任务没进展，额度却可能继续掉，难免让人着急。 这次OpenAI宣布改进插件和工具的使用能力，效果如何还要看「疗效」。 如果工具需要访问外部服务，用户仍要先连接账号、完成授权，AI只能在获准的权限范围内操作。 最核心的变化，是多个智能体并行干活。 比如，你想从三款AI工具里挑一款，需要比较功能、价格和使用限制。 AI可以分头搜集各款产品的资料，再把月付、年付、免费版和付费版的条件对齐。 用户@rubenssoto_a也分享了自己的尝试： 我让它做研究，再输出产品需求文档和任务工单，结果很棒。 从查资料到写需求、列任务，研究成果可以继续转成下一步的执行计划。 不过，AI分工和人类团队一样，能同时做的分头做，前后有依赖的则要按顺序来。 比如，先读材料、再定方案、最后修改文件，后一步要用到前一步的结果，多派几个AI也不能跳过。 拆出真正能同时推进的部分，才有机会缩短等待。 活儿分出去，进度也要能看得见。 用户可以在主对话里跟踪各个智能体，完成后再点开检查成果。 网友@wushanyunyu1提到这个变化时说，「这就从聊天变成了派活。」 在他看来，接下来要比的，是谁能把多个智能体的工作管好，少出错、少返工。 他更想看到的是，Pro用户用它做完一批真实任务后，到底有多少需要返工。 几个AI同时交活，人也可能看不过来。 网友@leolaoshuai提醒：「卡脖子的变成你验收的速度。」 如果每份成果都要从头检查，省下的等待时间，就可能花在验收上。 他的建议是，派活前写清楚「怎样算做完」，交付时附上能直接核对的依据。 智能体交出的成果，不仅要能点开查看，最好还要列清楚做了什么、依据在哪，让用户少花时间翻查和核对。 至于最多能同时跑几个智能体、分工是否默认触发、结果具体怎样合并，这条公告中还没有展开说明。 ## Claude早就组起了智能体军团 网友@FinallySomeRest在评论区说：「这个，Claude早就有了。」 说起让AI组队，Claude其实早有实践。 2025年6月公开的Claude Research，就能由主智能体带队，让多个子智能体分头研究，再把发现汇总回来。 主智能体分派研究任务，多个子智能体并行搜索，再汇总发现。（图源：Anthropic） 到了2026年9月17日，新版Claude Code Projects开始公测，把协作延伸到持续推进的项目中。 你在项目对话里交代目标，Claude负责拆任务、协调多条线程，并积累项目记忆。云端线程还能在你合上电脑后继续干活。 10月10日，官方又宣布放行Projects等待名单中的Pro和Max用户。 同期，Managed Agents动态工作流开启公测，并公布了一组查Bug测试： 11.6万行代码中预埋70个Bug，单个智能体跑三次，分别找出14、15和27个；换成动态工作流，三次都找出了66个。 从分头查资料，到协调多条任务线，Claude和ChatGPT都在接手更多分工与协作。 至于「谁来干、先干什么、结果怎么接上」，各家的实现方式和公开细节还不相同。 AI能自己分活、接好结果，人才能少当几次传话员。 ## 分头干活最后还得对得上 但是，把任务拆开，只是开始。 如果几个AI都去翻同一份资料，看起来忙得热火朝天，实际却在重复劳动。 如果一条支线用了最新数据，另一条还在参考旧版条件，拼出来的报告也可能前后打架。 Anthropic在介绍多智能体研究系统时，就分享过一次教训： 主智能体交代得不够具体，两个子智能体重复调查同一个方向，另一个却查起了不同年份的问题。 因此，分工时得说清楚：各自查什么、查到什么程度、按什么格式交结果。任务边界越含糊，越容易重复搜索，或者漏掉关键内容。 分头查完，接下来还得把结果对齐。 网友@878967502Teamo直言：「并行本身不难，难的是最后合并。」 他希望除了看进度，还能一眼看出几条支线的结果在哪里有冲突。 写报告，数据时间、价格条件和比较口径得一致；写代码，接口和命名要能对得上。 否则，AI分头干完了，人却得接着查错、补漏、重新拼。 Anthropic在Claude Code官方说明中也提醒，前后紧密依赖的步骤不适合强行拆开，多个智能体同时修改同一文件容易产生冲突。 任务太小，光是分工、交接，就可能抵消并行省下的时间。 ## 功能很酷 ## 但也要看清额度 能不能用上、能用多少，是用户接下来要算的账。 根据官方帮助说明，Plus用户可以在Work和Codex中使用Astra，但不能在Chat里使用GPT-6 Pro。 这次Chat中的多智能体升级，目前正向符合条件的Pro用户逐步推送，Plus用户暂时还用不上。 网友@WhosCinq就在评论区为Plus用户抱不平，质疑这次更新又把他们排除在外。 已经付费的用户，也在关心价格。 网友@JcCastaneda05晒出自己的Pro订阅和用量页面，诉求就一句话： 请让它更便宜。 开通Pro，也不意味着可以放开跑—— 先看在哪儿用。 Chat里的Pro模型消息限额，与Work、Codex共享的用量池分开计算。 上面这张用户截图也写明，共享用量不包含Chat对话。 然后，再看订阅哪一档。 当前官方帮助文档显示，200美元档Pro的GPT-6 Pro有周限额；100美元档Pro中，GPT-6 Pro和GPT-5.6 Sol Pro共用一份周额度。 200美元档GPT-6 Pro有周限额；100美元档两款Pro模型共用周额度。（图源：OpenAI） 除了额度够不够，用户还希望每一笔消耗都清楚。 网友@aiertixue称，自己的Codex周额度突然从剩余约60%降到了0%，却没有相应的使用，希望团队调查并恢复额度。 这条反馈说的是Codex，不能据此判断Chat里的多智能体怎样扣额度。 这背后是一个实际需求：任务还没做完，用户得知道额度花在了哪里。 ## AI交卷之后 ## 人还要填多少坑 一支AI小队好不好用，要看从接到任务到交出合格成果，总共花了多少时间、消耗了多少额度，以及还需要人做多少检查和修改。 用户@cooolernemesis就提出了一个具体的评测要求： 把协调开销和重复工具调用也算上，在相同质量要求下，整项工作究竟快了多少？ 毕竟，AI跑得快，不代表用户就能早点收工。 把同一份资料、同一个复杂需求交给ChatGPT和Claude，先看谁遗漏的信息少，谁的结论有依据，再看用户还要补几次要求、花多久返工。 跟踪进度，能看出哪条支线卡住了；打开已完成的任务，能检查资料有没有遗漏、代码有没有写错。 这些设计能否省事，最终还得看它们能不能帮助用户更快发现问题、拿到合格成果。 AI交卷，不等于工作结束。 人查完、改完，成果真正能用上，整项任务才算完成。 参考资料： https://x.com/OpenAI/status/2111192259073094035 https://help.openai.com/en/articles/6825453-chatgpt-release-notes https://www.anthropic.com/engineering/multi-agent-research-system https://claude.com/resources/articles/subagents-in-claude-cod 本文来自微信公众号“新智元”（ID：AI_era），作者：ASI启示录，编辑：元宇,36氪经授权发布。 | https://cdn.deepseek.com/site-icons/36kr.com | |
| 6 | Interesting Engineering | OpenAI models hit 97.9% accuracy in tests to modernize old software | https://interestingengineering.com/ai-robotics/openai-coding-models-software-modernization#main-content | 2.24 | 3 | OpenAI’s GPT-6 Astra scored 87.4 percent on a coding benchmark spanning 20 programming languages, up from 26.8 percent. ... GPT-6 Astra achieved an average test pass rate of 87.4 percent, up from 26.8 percent for GPT-5.6 Sol. | # OpenAI models make huge gains in modernizing decades-old software with 97.9% accuracy OpenAI’s GPT-6 Astra scored 87.4 percent on a coding benchmark spanning 20 programming lang…Show full text# OpenAI models make huge gains in modernizing decades-old software with 97.9% accuracy OpenAI’s GPT-6 Astra scored 87.4 percent on a coding benchmark spanning 20 programming languages, up from 26.8 percent. ... OpenAI has significantly improved its models’ ability to write software in older programming languages and migrate existing code. Its latest results include 97.9 percent accuracy on a test designed to measure the migration of legacy programs. ... OpenAI also tested its models on MultiLangBench, which measures their ability to build a database engine compatible with SQLite across 20 programming languages. GPT-6 Astra achieved an average test pass rate of 87.4 percent, up from 26.8 percent for GPT-5.6 Sol. Astra improved in every language tested. Its Fortran score climbed from 21.4 percent to 83.5 percent, while Visual Basic rose from 27.5 percent to 89.4 percent. COBOL remained the hardest language, although Astra raised its score from 10.9 percent to 72.1 percent. ... Astra’s results varied considerably across languages, and COBOL still posed the greatest challenge. ... The results offer encouraging evidence of progress, but teams will need to carefully validate production deployments before trusting migrated software with critical operations. | # OpenAI models make huge gains in modernizing decades-old software with 97.9% accuracy OpenAI’s GPT-6 Astra scored 87.4 percent on a coding benchmark spanning 20 programming lang…Show full text# OpenAI models make huge gains in modernizing decades-old software with 97.9% accuracy OpenAI’s GPT-6 Astra scored 87.4 percent on a coding benchmark spanning 20 programming languages, up from 26.8 percent. Read Next: Chinese think tank calls for faster, more rigorous assessment of US defense technologies OpenAI has significantly improved its models’ ability to write software in older programming languages and migrate existing code. Its latest results include 97.9 percent accuracy on a test designed to measure the migration of legacy programs. The improvements could help businesses modernize software that remains essential to daily operations. Banks and financial institutions still rely on decades-old code, making replacement risky when developers must preserve established calculations and business rules. ## Tackling legacy code COBOL remains deeply embedded in financial infrastructure. IBM estimates that it supports more than 40 percent of online banking systems, 80 percent of in-person credit card transactions, and 95 percent of ATM transactions. Fortran also remains important in scientific computing, including weather forecasting and nuclear reactor safety analysis. Both languages have less training material available than popular alternatives such as Python and JavaScript. OpenAI said developers and enterprise customers had reported difficulties getting coding models to work reliably with these older languages. Documentation can be incomplete, and essential business rules may exist only in the original code. That creates problems during migration. A rewritten program might run without errors yet produce different results. COBOL, for example, can discard excess decimal digits unless a program explicitly requests rounding. Changing that behavior could affect financial calculations. ### More from AI and Robotics See All ## Rewriting code accurately OpenAI’s COBOL migration evaluation tests whether models can translate existing programs into C++, Java, Kotlin, Python, or Rust without changing their behavior. GPT-6.1 Sol achieved 86.6 percent accuracy at low reasoning effort, compared with 5.4 percent for GPT-5.6 Sol. OpenAI also reported a 39 percent reduction in API costs at that setting. At higher reasoning effort, GPT-6.1 Sol reached 97.9 percent accuracy. The evaluation requires translated programs to pass every hidden test, including checks that compare their observable outputs with the original programs. That figure reflects performance on OpenAI’s test suite, not a guarantee of error-free migrations. Developers must still verify rewritten applications against real workloads and undocumented dependencies. ## Building across languages OpenAI also tested its models on MultiLangBench, which measures their ability to build a database engine compatible with SQLite across 20 programming languages. GPT-6 Astra achieved an average test pass rate of 87.4 percent, up from 26.8 percent for GPT-5.6 Sol. Astra improved in every language tested. Its Fortran score climbed from 21.4 percent to 83.5 percent, while Visual Basic rose from 27.5 percent to 89.4 percent. COBOL remained the hardest language, although Astra raised its score from 10.9 percent to 72.1 percent. GPT-6 Luna delivered a lower average pass rate of 40.3 percent but cost approximately 78 percent less per sample than GPT-5.6 Sol. OpenAI estimated average API costs of $0.77 and $3.48, respectively. The findings point to stronger support for maintaining aging software. However, developers must still ensure that modernized systems preserve the behavior businesses depend on. The benchmarks also show where more work is needed. Astra’s results varied considerably across languages, and COBOL still posed the greatest challenge. Meanwhile, Luna’s lower costs could let engineering teams run more experiments without substantially increasing their budgets. The results offer encouraging evidence of progress, but teams will need to carefully validate production deployments before trusting migrated software with critical operations. Get the latest in engineering, tech, space & science - delivered daily to your inbox. By subscribing, you agree to our Terms of Use and Policies You may unsubscribe at any time. 0COMMENT Subscribe to Today! Access to exclusive content, expert insights and a deeper dive into engineering and tech. No ads, no limits. Explore Now! Aamir is a seasoned tech journalist with experience at Exhibit Magazine, Republic World, and PR Newswire. With a deep love for all things tech and science, he has spent years decoding the latest innovations and exploring how they shape industries, lifestyles, and the future of humanity. ADVERTISEMENT | https://cdn.deepseek.com/site-icons/interestingengineering.com | |
| 7 | Amazon AWS Documentation | GPT-6 Astra - GPT-6 Astra | https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-6-astra.html#1 | 2.20 | 4 | Pricing All prices are in USD per 1 million tokens. The following tables list Standard and Ultrafast prices. The Input — 30-minute ... Inference option | Input | Input — 30-minute …Show full textPricing All prices are in USD per 1 million tokens. The following tables list Standard and Ultrafast prices. The Input — 30-minute ... Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ... Ultrafast is a speed tier for GPT-6 Astra. | ###### Note ... In outputConfig.textFormat.structure.jsonSchema, set name and schema. Encode the schema as a JSON string. Then set additionalModelRequestFields.text.format.strict t…Show full text###### Note ... In outputConfig.textFormat.structure.jsonSchema, set name and schema. Encode the schema as a JSON string. Then set additionalModelRequestFields.text.format.strict to true. Use an object schema. Put all fields inrequired. Set additionalProperties to false. Validate your schema before you send it. Check for refusals or incomplete responses before you parse the output. ###### Important For JSON Schema output with the Converse API, you must setadditionalModelRequestFields.text.format.strict to true, in addition to specifying the schema in outputConfig. Include the following field in your request. { "additionalModelRequestFields": ... { "strict": ... ## Pricing All prices are in USD per 1 million tokens. The following tables list Standard and Ultrafast prices. The Input — 30-minute cache write column is the price per 1 million tokens written to the prompt cache. The 30-minute figure is how long a cached prompt prefix is retained, not a quantity of tokens. For more information, see prompt caching. Commercial In-Region and Geo CRIS prices include a 10% premium over the corresponding OpenAI rates for the same service tier. You do not need to add this premium. Priority and Flex tiers are not supported for this model. ### Standard — Commercial Regions, short context (272K input tokens or fewer) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region | $11.00 | $13.75 | $1.10 | $55.00 Geo CRIS | $11.00 | $13.75 | $1.10 | $55.00 Global CRIS | $10.00 | $12.50 | $1.00 | $50.00 ### Standard — Commercial Regions, long context (more than 272K input tokens) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region | $22.00 | $27.50 | $2.20 | $82.50 Geo CRIS | $22.00 | $27.50 | $2.20 | $82.50 Global CRIS | $20.00 | $25.00 | $2.00 | $75.00 Ultrafast prices are six times the corresponding Standard prices. Onbedrock-runtime, Ultrafast is available through both US geographic CRIS and Global CRIS. On bedrock-mantle, Ultrafast is available in us-east-1. Short-context prices apply to requests with at most 272,000 input tokens. When input exceeds this threshold, long-context prices apply to the entire request. ### Ultrafast — Commercial Regions, short context (272K input tokens or fewer) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region (us-east-1) | $66.00 | $82.50 | $6.60 | $330.00 Geo CRIS (US) | $66.00 | $82.50 | $6.60 | $330.00 Global CRIS | $60.00 | $75.00 | $6.00 | $300.00 ### Ultrafast — Commercial Regions, long context (more than 272K input tokens) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region (us-east-1) | $132.00 | $165.00 | $13.20 | $495.00 Geo CRIS (US) | $132.00 | $165.00 | $13.20 | $495.00 Global CRIS | $120.00 | $150.00 | $12.00 | $450.00 ## Programmatic Access To call this model from code, use the following model IDs and endpoint URLs. ... Endpoint | Model ID | In-... | # GPT-6 Astra ## OpenAI — GPT-6 Astra ## Model Details GPT-6 Astra is OpenAI's most capable model. It handles the hardest end-to-end work. Use it for complex reasoning, coding, …Show full text# GPT-6 Astra ## OpenAI — GPT-6 Astra ## Model Details GPT-6 Astra is OpenAI's most capable model. It handles the hardest end-to-end work. Use it for complex reasoning, coding, computer use, research, and document creation. To learn more, see the OpenAI Deployment Safety Hub Model launch date: September 8, 2026 EOL no sooner than: September 8, 2027 Legacy period: at least 6 months Model lifecycle policy: Model lifecycle Model EOL date: N/A End User License Agreements and Terms of Use: View Model lifecycle: Active Context window: 1,050,000 tokens Max output tokens: 128,000 Knowledge cutoff: April 30, 2026 Marketplace product ID: prod-hqau7gqhlqsrg Input Modalities | Output Modalities ---|--- Audio | Embedding Image | Image Speech | Speech Text | Text Video | Video ## Endpoints and APIs supported The following tables show which endpoints and APIs GPT-6 Astra supports. For more information, see APIs supported by Amazon Bedrock and Inference endpoints supported by Amazon Bedrock. Endpoint support Endpoint | Supported ---|--- bedrock-runtime bedrock-mantle APIs supported onbedrock-runtime Messages | Responses | Chat Completions | Converse | Invoke APIs supported onbedrock-mantle Messages | Responses | Chat Completions | Converse | Invoke ###### Note Onbedrock-mantle, both APIs use the /openai/v1 base path, not /v1. Use either API with this model: For Responses, use /openai/v1/responses. For Chat Completions, use /openai/v1/chat/completions. ## Capabilities and Features Bedrock Features Features supported usingbedrock-runtime Features supported usingbedrock-mantle Supported | Not Supported ---|--- Implicit Prompt Caching (Responses API only) Explicit Prompt Caching (Responses API only) Reasoning Summaries (Responses API only) ### JSON Schema output on bedrock-runtime Use JSON Schema to set the format of the model response. These settings apply to non-streaming calls on thebedrock-runtime endpoint. Choose a profile ID from Programmatic access. For Chat Completions, set response_format.type to json_schema. Put name, schema, and strict: true in response_format.json_schema. For Responses, set text.format.type to json_schema. Put name, schema, and strict: true in text.format. For Converse, set outputConfig.textFormat.type to json_schema. In outputConfig.textFormat.structure.jsonSchema, set name and schema. Encode the schema as a JSON string. Then set additionalModelRequestFields.text.format.strict to true. Use an object schema. Put all fields inrequired. Set additionalProperties to false. Validate your schema before you send it. Check for refusals or incomplete responses before you parse the output. ###### Important For JSON Schema output with the Converse API, you must setadditionalModelRequestFields.text.format.strict to true, in addition to specifying the schema in outputConfig. Include the following field in your request. { "additionalModelRequestFields": { "text": { "format": { "strict": true } } } } ## Pricing All prices are in USD per 1 million tokens. The following tables list Standard and Ultrafast prices. The Input — 30-minute cache write column is the price per 1 million tokens written to the prompt cache. The 30-minute figure is how long a cached prompt prefix is retained, not a quantity of tokens. For more information, see prompt caching. Commercial In-Region and Geo CRIS prices include a 10% premium over the corresponding OpenAI rates for the same service tier. You do not need to add this premium. Priority and Flex tiers are not supported for this model. ### Standard — Commercial Regions, short context (272K input tokens or fewer) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region | $11.00 | $13.75 | $1.10 | $55.00 Geo CRIS | $11.00 | $13.75 | $1.10 | $55.00 Global CRIS | $10.00 | $12.50 | $1.00 | $50.00 ### Standard — Commercial Regions, long context (more than 272K input tokens) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region | $22.00 | $27.50 | $2.20 | $82.50 Geo CRIS | $22.00 | $27.50 | $2.20 | $82.50 Global CRIS | $20.00 | $25.00 | $2.00 | $75.00 Ultrafast prices are six times the corresponding Standard prices. Onbedrock-runtime, Ultrafast is available through both US geographic CRIS and Global CRIS. On bedrock-mantle, Ultrafast is available in us-east-1. Short-context prices apply to requests with at most 272,000 input tokens. When input exceeds this threshold, long-context prices apply to the entire request. ### Ultrafast — Commercial Regions, short context (272K input tokens or fewer) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region (us-east-1) | $66.00 | $82.50 | $6.60 | $330.00 Geo CRIS (US) | $66.00 | $82.50 | $6.60 | $330.00 Global CRIS | $60.00 | $75.00 | $6.00 | $300.00 ### Ultrafast — Commercial Regions, long context (more than 272K input tokens) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region (us-east-1) | $132.00 | $165.00 | $13.20 | $495.00 Geo CRIS (US) | $132.00 | $165.00 | $13.20 | $495.00 Global CRIS | $120.00 | $150.00 | $12.00 | $450.00 ## Programmatic Access To call this model from code, use the following model IDs and endpoint URLs. For more information, see APIs supported by Amazon Bedrock and Inference endpoints supported by Amazon Bedrock. Endpoint | Model ID | In-Region endpoint URL | Geo inference ID | Global inference ID ---|---|---|---|--- bedrock-mantle | openai.gpt-6-astra | https://bedrock-mantle.us-east-1.api.aws/openai/v1 https://bedrock-mantle.us-west-2.api.aws/openai/v1 (Standard only) | Not supported | Not supported bedrock-runtime | openai.gpt-6-astra | Not supported | us.openai.gpt-6-astra | global.openai.gpt-6-astra Thebedrock-mantle endpoint is available in us-east-1 (N. Virginia) and us-west-2 (Oregon). Ultrafast is available only in us-east-1 on this endpoint. On bedrock-runtime, the base URL is https://bedrock-runtime.{region}.amazonaws.com/openai/v1. Name us.openai.gpt-6-astra for US geographic cross-Region inference or global.openai.gpt-6-astra for global cross-Region inference. ## Service Tiers Amazon Bedrock offers several service tiers for different workloads. Standard gives you pay-per-token access with no commitment. To use it, set"service_tier": "default" or omit the field. For more information, see service tiers. Ultrafast is a speed tier for GPT-6 Astra. OpenAI reports API speeds up to six times faster than Standard. To request it with the Responses API, set"service_tier": "ultrafast". Standard | Ultrafast | Priority | Flex | Reserved Use one of these routes for Ultrafast: - bedrock-mantle in us-east-1, with model ID openai.gpt-6-astra. - bedrock-runtime, with us.openai.gpt-6-astra for US geographic CRIS or global.openai.gpt-6-astra for Global CRIS. ###### Note Ultrafast does not support regional Mantle access inus-west-2. If an existing session is pinned to Oregon capacity, start a new session using an Ultrafast-supported route. ## Regional Availability Regional availability at a glance Amazon Bedrock offers three inference options. In-Region keeps requests in one Region for strict compliance. Geo Cross-Region routes requests across Regions in one geography. It respects data residency. Global Cross-Region routes requests anywhere in the world. Use it when you have no data residency needs. For more information, see the Regional availability by models page. Availability using thebedrock-mantle endpoint Region | In-Region | Geo | Global ---|---|---|--- us-east-1 (N. Virginia) us-west-2 (Oregon) Ultrafast is available through regional Mantle inus-east-1. Regional Mantle in us-west-2 supports Standard only. Availability using thebedrock-runtime endpoint Region | In-Region | Geo | Global ---|---|---|--- us-east-1 (N. Virginia) us-east-2 (Ohio) us-west-1 (N. California) us-west-2 (Oregon) ca-central-1 (Canada) ca-west-1 (Calgary) eu-central-1 (Frankfurt) eu-north-1 (Stockholm) eu-west-1 (Ireland) eu-west-2 (London) eu-west-3 (Paris) ap-northeast-1 (Tokyo) ap-northeast-2 (Seoul) ap-northeast-3 (Osaka) ap-south-1 (Mumbai) ap-southeast-1 (Singapore) ap-southeast-2 (Sydney) sa-east-1 (São Paulo) Geo inference details The destination Regions available to a geographic inference profile depend on the source Region. To retrieve the current routing configuration, call GetInferenceProfile from the source Region. Geo: US Geo inference ID:us.openai.gpt-6-astra Source Region | Destination Regions ---|--- us-east-1 (N. Virginia) | us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) us-east-2 (Ohio) | us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) us-west-1 (N. California) | us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-1 (N. California), us-west-2 (Oregon) us-west-2 (Oregon) | us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) ca-central-1 (Canada) | ca-central-1 (Canada), us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) ca-west-1 (Calgary) | ca-west-1 (Calgary), us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) ## Quotas and Limits Your AWS account has default quotas. These quotas help keep Amazon Bedrock running well. A few things can change your quotas. These include Region, payment history, fraudulent use, or an approved quota increase request. For more information, see the Quotas for Amazon Bedrock documentation and the limits for the model. On thebedrock-runtime endpoint, limits are managed as tokens per minute (TPM) with a 10x burndown rate, where 1 output token consumes 10 tokens. ## Sample Code Step 1 - AWS Account: If you already have an AWS account, skip this step. If you are new to AWS, sign up for an AWS account Step 2 - API key: Go to the Amazon Bedrock console Step 3 - Get the SDK: You must have Python installed to use this guide. Then install the OpenAI SDK. pip install openai ### Step 4 - Set environment variables Set up your environment to use the API key for authentication. - bedrock-mantle - OPENAI_API_KEY="<provide your Bedrock API key>" OPENAI_BASE_URL="https://bedrock-mantle.us-west-2.api.aws/openai/v1" - bedrock-runtime - OPENAI_API_KEY="<provide your Bedrock API key>" OPENAI_BASE_URL="https://bedrock-runtime.us-west-2.amazonaws.com/openai/v1" ###### Note Onbedrock-runtime, name a cross-Region inference profile as the model: us.openai.gpt-6-astra or global.openai.gpt-6-astra. This model is not available for in-Region inference on that endpoint. ### Step 5 - Run your first inference request Save the file asbedrock-first-request.py. #### bedrock-mantle Use the settings from Step 4 - Set environment variables. Choose thebedrock-mantle tab. The Chat tab uses the Chat Completions API. - Responses API - from openai import OpenAI client = OpenAI() response = client.responses.create( model="openai.gpt-6-astra", input="Can you explain the features of Amazon Bedrock?" ) print(response) - Chat - from openai import OpenAI client = OpenAI() response = client.chat.completions.create( model="openai.gpt-6-astra", messages=[{"role": "user", "content": "Can you explain the features of Amazon Bedrock?"}] ) print(response) #### bedrock-runtime: OpenAI SDK Use the settings from Step 4 - Set environment variables. Choose thebedrock-runtime tab. Send your request with the Responses API. - Responses API - from openai import OpenAI client = OpenAI() response = client.responses.create( model="us.openai.gpt-6-astra", input="Can you explain the features of Amazon Bedrock?" ) print(response) ## Ultrafast example SetOPENAI_API_KEY to your Amazon Bedrock API key. This example uses the bedrock-mantle endpoint in us-east-1. export OPENAI_API_KEY="<your Amazon Bedrock API key>"from openai import OpenAI client = OpenAI( base_url="https://bedrock-mantle.us-east-1.api.aws/openai/v1", ) response = client.responses.create( model="openai.gpt-6-astra", service_tier="ultrafast", input="Explain how Amazon Bedrock cross-Region inference works.", ) print(response.output_text) Forbedrock-runtime, use the OpenAI-compatible base URL from Programmatic Access. Set the model to us.openai.gpt-6-astra for US geographic CRIS or global.openai.gpt-6-astra for Global CRIS, and keep service_tier="ultrafast". Did this page help you? - Yes Thanks for letting us know we're doing a good job! If you've got a moment, please tell us what we did right so we can do more of it. Did this page help you? - No Thanks for letting us know this page needs work. We're sorry we let you down. If you've got a moment, please tell us how we can make the documentation better. | https://cdn.deepseek.com/site-icons/amazon.com | Unknown |
| 8 | LaunchTemplateInstanceMetadataOptionsRequest | GPT-6 Astra | https://docs.aws.eu//bedrock/latest/userguide/model-card-openai-gpt-6-astra.html | 2.20 | 4 | All prices are in USD per 1 million tokens. The following tables list Standard and Ultrafast prices. The Input — 30-minute ... Inference option | Input | Input — 30-minute cache wr…Show full textAll prices are in USD per 1 million tokens. The following tables list Standard and Ultrafast prices. The Input — 30-minute ... Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ... Ultrafast is a speed tier for GPT-6 Astra. | ###### Note ... Set additionalProperties to false. Validate your schema before you send it. Check for refusals or incomplete responses before you parse the output. ###### Importan…Show full text###### Note ... Set additionalProperties to false. Validate your schema before you send it. Check for refusals or incomplete responses before you parse the output. ###### Important For JSON Schema output with the Converse API, you must setadditionalModelRequestFields.text.format.strict to true, in addition to specifying the schema in outputConfig. Include the following field in your request. { "additionalModelRequestFields": ... { "strict": ... ## Pricing All prices are in USD per 1 million tokens. The following tables list Standard and Ultrafast prices. The Input — 30-minute cache write column is the price per 1 million tokens written to the prompt cache. The 30-minute figure is how long a cached prompt prefix is retained, not a quantity of tokens. For more information, see prompt caching. Commercial In-Region and Geo CRIS prices include a 10% premium over the corresponding OpenAI rates for the same service tier. You do not need to add this premium. Priority and Flex tiers are not supported for this model. ### Standard — Commercial Regions, short context (272K input tokens or fewer) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region | $11.00 | $13.75 | $1.10 | $55.00 Geo CRIS | $11.00 | $13.75 | $1.10 | $55.00 Global CRIS | $10.00 | $12.50 | $1.00 | $50.00 ### Standard — Commercial Regions, long context (more than 272K input tokens) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region | $22.00 | $27.50 | $2.20 | $82.50 Geo CRIS | $22.00 | $27.50 | $2.20 | $82.50 Global CRIS | $20.00 | $25.00 | $2.00 | $75.00 Ultrafast prices are six times the corresponding Standard prices. Onbedrock-runtime, Ultrafast is available through both US geographic CRIS and Global CRIS. On bedrock-mantle, Ultrafast is available in us-east-1. Short-context prices apply to requests with at most 272,000 input tokens. When input exceeds this threshold, long-context prices apply to the entire request. ### Ultrafast — Commercial Regions, short context (272K input tokens or fewer) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region (us-east-1) | $66.00 | $82.50 | $6.60 | $330.00 Geo CRIS (US) | $66.00 | $82.50 | $6.60 | $330.00 Global CRIS | $60.00 | $75.00 | $6.00 | $300.00 ### Ultrafast — Commercial Regions, long context (more than 272K input tokens) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region (us-east-1) | $132.00 | $165.00 | $13.20 | $495.00 Geo CRIS (US) | $132.00 | $165.00 | $13.20 | $495.00 Global CRIS | $120.00 | $150.00 | $12.00 | $450.00 ## Programmatic Access To call this model from code, use the following model IDs and endpoint URLs. ... Endpoint | Model ID | In-Region endpoint URL | Geo inference ID | Global inference ID ---|---|---|---|--- bedrock-mantle | openai.gpt-6-astra | https://bedrock-mantle.us-east-1.api.aws/openai... | - Services or capabilities described in AWS documentation might vary by Region. To see the differences applicable to the AWS European Sovereign Cloud Region, see the AWS European S…Show full text- Services or capabilities described in AWS documentation might vary by Region. To see the differences applicable to the AWS European Sovereign Cloud Region, see the AWS European Sovereign Cloud User Guide. ## OpenAI — GPT-6 Astra ## Model Details GPT-6 Astra is OpenAI's most capable model. It handles the hardest end-to-end work. Use it for complex reasoning, coding, computer use, research, and document creation. To learn more, see the OpenAI Deployment Safety Hub Model launch date: September 8, 2026 EOL no sooner than: September 8, 2027 Legacy period: at least 6 months Model lifecycle policy: Model lifecycle Model EOL date: N/A End User License Agreements and Terms of Use: View Model lifecycle: Active Context window: 1,050,000 tokens Max output tokens: 128,000 Knowledge cutoff: April 30, 2026 Marketplace product ID: prod-hqau7gqhlqsrg Input Modalities | Output Modalities ---|--- Audio | Embedding Image | Image Speech | Speech Text | Text Video | Video ## Endpoints and APIs supported The following tables show which endpoints and APIs GPT-6 Astra supports. For more information, see APIs supported by Amazon Bedrock and Inference endpoints supported by Amazon Bedrock. Endpoint support Endpoint | Supported ---|--- bedrock-runtime bedrock-mantle APIs supported onbedrock-runtime Messages | Responses | Chat Completions | Converse | Invoke APIs supported onbedrock-mantle Messages | Responses | Chat Completions | Converse | Invoke ###### Note Onbedrock-mantle, both APIs use the /openai/v1 base path, not /v1. Use either API with this model: For Responses, use /openai/v1/responses. For Chat Completions, use /openai/v1/chat/completions. ## Capabilities and Features Bedrock Features Features supported usingbedrock-runtime Features supported usingbedrock-mantle Supported | Not Supported ---|--- Implicit Prompt Caching (Responses API only) Explicit Prompt Caching (Responses API only) Reasoning Summaries (Responses API only) ### JSON Schema output on bedrock-runtime Use JSON Schema to set the format of the model response. These settings apply to non-streaming calls on thebedrock-runtime endpoint. Choose a profile ID from Programmatic access. For Chat Completions, set response_format.type to json_schema. Put name, schema, and strict: true in response_format.json_schema. For Responses, set text.format.type to json_schema. Put name, schema, and strict: true in text.format. For Converse, set outputConfig.textFormat.type to json_schema. In outputConfig.textFormat.structure.jsonSchema, set name and schema. Encode the schema as a JSON string. Then set additionalModelRequestFields.text.format.strict to true. Use an object schema. Put all fields inrequired. Set additionalProperties to false. Validate your schema before you send it. Check for refusals or incomplete responses before you parse the output. ###### Important For JSON Schema output with the Converse API, you must setadditionalModelRequestFields.text.format.strict to true, in addition to specifying the schema in outputConfig. Include the following field in your request. { "additionalModelRequestFields": { "text": { "format": { "strict": true } } } } ## Pricing All prices are in USD per 1 million tokens. The following tables list Standard and Ultrafast prices. The Input — 30-minute cache write column is the price per 1 million tokens written to the prompt cache. The 30-minute figure is how long a cached prompt prefix is retained, not a quantity of tokens. For more information, see prompt caching. Commercial In-Region and Geo CRIS prices include a 10% premium over the corresponding OpenAI rates for the same service tier. You do not need to add this premium. Priority and Flex tiers are not supported for this model. ### Standard — Commercial Regions, short context (272K input tokens or fewer) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region | $11.00 | $13.75 | $1.10 | $55.00 Geo CRIS | $11.00 | $13.75 | $1.10 | $55.00 Global CRIS | $10.00 | $12.50 | $1.00 | $50.00 ### Standard — Commercial Regions, long context (more than 272K input tokens) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region | $22.00 | $27.50 | $2.20 | $82.50 Geo CRIS | $22.00 | $27.50 | $2.20 | $82.50 Global CRIS | $20.00 | $25.00 | $2.00 | $75.00 Ultrafast prices are six times the corresponding Standard prices. Onbedrock-runtime, Ultrafast is available through both US geographic CRIS and Global CRIS. On bedrock-mantle, Ultrafast is available in us-east-1. Short-context prices apply to requests with at most 272,000 input tokens. When input exceeds this threshold, long-context prices apply to the entire request. ### Ultrafast — Commercial Regions, short context (272K input tokens or fewer) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region (us-east-1) | $66.00 | $82.50 | $6.60 | $330.00 Geo CRIS (US) | $66.00 | $82.50 | $6.60 | $330.00 Global CRIS | $60.00 | $75.00 | $6.00 | $300.00 ### Ultrafast — Commercial Regions, long context (more than 272K input tokens) Inference option | Input | Input — 30-minute cache write | Input — cache read | Output ---|---|---|---|--- In-Region (us-east-1) | $132.00 | $165.00 | $13.20 | $495.00 Geo CRIS (US) | $132.00 | $165.00 | $13.20 | $495.00 Global CRIS | $120.00 | $150.00 | $12.00 | $450.00 ## Programmatic Access To call this model from code, use the following model IDs and endpoint URLs. For more information, see APIs supported by Amazon Bedrock and Inference endpoints supported by Amazon Bedrock. Endpoint | Model ID | In-Region endpoint URL | Geo inference ID | Global inference ID ---|---|---|---|--- bedrock-mantle | openai.gpt-6-astra | https://bedrock-mantle.us-east-1.api.aws/openai/v1 https://bedrock-mantle.us-west-2.api.aws/openai/v1 (Standard only) | Not supported | Not supported bedrock-runtime | openai.gpt-6-astra | Not supported | us.openai.gpt-6-astra | global.openai.gpt-6-astra Thebedrock-mantle endpoint is available in us-east-1 (N. Virginia) and us-west-2 (Oregon). Ultrafast is available only in us-east-1 on this endpoint. On bedrock-runtime, the base URL is https://bedrock-runtime.{region}.amazonaws.com/openai/v1. Name us.openai.gpt-6-astra for US geographic cross-Region inference or global.openai.gpt-6-astra for global cross-Region inference. ## Service Tiers Amazon Bedrock offers several service tiers for different workloads. Standard gives you pay-per-token access with no commitment. To use it, set"service_tier": "default" or omit the field. For more information, see service tiers. Ultrafast is a speed tier for GPT-6 Astra. OpenAI reports API speeds up to six times faster than Standard. To request it with the Responses API, set"service_tier": "ultrafast". Standard | Ultrafast | Priority | Flex | Reserved Use one of these routes for Ultrafast: - bedrock-mantle in us-east-1, with model ID openai.gpt-6-astra. - bedrock-runtime, with us.openai.gpt-6-astra for US geographic CRIS or global.openai.gpt-6-astra for Global CRIS. ###### Note Ultrafast does not support regional Mantle access inus-west-2. If an existing session is pinned to Oregon capacity, start a new session using an Ultrafast-supported route. ## Regional Availability Regional availability at a glance Amazon Bedrock offers three inference options. In-Region keeps requests in one Region for strict compliance. Geo Cross-Region routes requests across Regions in one geography. It respects data residency. Global Cross-Region routes requests anywhere in the world. Use it when you have no data residency needs. For more information, see the Regional availability by models page. Availability using thebedrock-mantle endpoint Region | In-Region | Geo | Global ---|---|---|--- us-east-1 (N. Virginia) us-west-2 (Oregon) Ultrafast is available through regional Mantle inus-east-1. Regional Mantle in us-west-2 supports Standard only. Availability using thebedrock-runtime endpoint Region | In-Region | Geo | Global ---|---|---|--- us-east-1 (N. Virginia) us-east-2 (Ohio) us-west-1 (N. California) us-west-2 (Oregon) ca-central-1 (Canada) ca-west-1 (Calgary) eu-central-1 (Frankfurt) eu-north-1 (Stockholm) eu-west-1 (Ireland) eu-west-2 (London) eu-west-3 (Paris) ap-northeast-1 (Tokyo) ap-northeast-2 (Seoul) ap-northeast-3 (Osaka) ap-south-1 (Mumbai) ap-southeast-1 (Singapore) ap-southeast-2 (Sydney) sa-east-1 (São Paulo) Geo inference details The destination Regions available to a geographic inference profile depend on the source Region. To retrieve the current routing configuration, call GetInferenceProfile from the source Region. Geo: US Geo inference ID:us.openai.gpt-6-astra Source Region | Destination Regions ---|--- us-east-1 (N. Virginia) | us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) us-east-2 (Ohio) | us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) us-west-1 (N. California) | us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-1 (N. California), us-west-2 (Oregon) us-west-2 (Oregon) | us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) ca-central-1 (Canada) | ca-central-1 (Canada), us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) ca-west-1 (Calgary) | ca-west-1 (Calgary), us-east-1 (N. Virginia), us-east-2 (Ohio), us-west-2 (Oregon) ## Quotas and Limits Your AWS account has default quotas. These quotas help keep Amazon Bedrock running well. A few things can change your quotas. These include Region, payment history, fraudulent use, or an approved quota increase request. For more information, see the Quotas for Amazon Bedrock documentation and the limits for the model. On thebedrock-runtime endpoint, limits are managed as tokens per minute (TPM) with a 10x burndown rate, where 1 output token consumes 10 tokens. ## Sample Code Step 1 - AWS Account: If you already have an AWS account, skip this step. If you are new to AWS, sign up for an AWS account Step 2 - API key: Go to the Amazon Bedrock console Step 3 - Get the SDK: You must have Python installed to use this guide. Then install the OpenAI SDK. pip install openai ### Step 4 - Set environment variables Set up your environment to use the API key for authentication. - bedrock-mantle - OPENAI_API_KEY="<provide your Bedrock API key>" OPENAI_BASE_URL="https://bedrock-mantle.us-west-2.api.aws/openai/v1" - bedrock-runtime - OPENAI_API_KEY="<provide your Bedrock API key>" OPENAI_BASE_URL="https://bedrock-runtime.us-west-2.amazonaws.com/openai/v1" OPENAI_API_KEY="<provide your Bedrock API key>" OPENAI_BASE_URL="https://bedrock-mantle.us-west-2.api.aws/openai/v1" ###### Note Onbedrock-runtime, name a cross-Region inference profile as the model: us.openai.gpt-6-astra or global.openai.gpt-6-astra. This model is not available for in-Region inference on that endpoint. ### Step 5 - Run your first inference request Save the file asbedrock-first-request.py. #### bedrock-mantle Use the settings from Step 4 - Set environment variables. Choose thebedrock-mantle tab. The Chat tab uses the Chat Completions API. - Responses API - from openai import OpenAI client = OpenAI() response = client.responses.create( model="openai.gpt-6-astra", input="Can you explain the features of Amazon Bedrock?" ) print(response) - Chat - from openai import OpenAI client = OpenAI() response = client.chat.completions.create( model="openai.gpt-6-astra", messages=[{"role": "user", "content": "Can you explain the features of Amazon Bedrock?"}] ) print(response) from openai import OpenAI client = OpenAI() response = client.responses.create( model="openai.gpt-6-astra", input="Can you explain the features of Amazon Bedrock?" ) print(response) #### bedrock-runtime: OpenAI SDK Use the settings from Step 4 - Set environment variables. Choose thebedrock-runtime tab. Send your request with the Responses API. - Responses API - from openai import OpenAI client = OpenAI() response = client.responses.create( model="us.openai.gpt-6-astra", input="Can you explain the features of Amazon Bedrock?" ) print(response) from openai import OpenAI client = OpenAI() response = client.responses.create( model="us.openai.gpt-6-astra", input="Can you explain the features of Amazon Bedrock?" ) print(response) ## Ultrafast example SetOPENAI_API_KEY to your Amazon Bedrock API key. This example uses the bedrock-mantle endpoint in us-east-1. export OPENAI_API_KEY="<your Amazon Bedrock API key>"from openai import OpenAI client = OpenAI( base_url="https://bedrock-mantle.us-east-1.api.aws/openai/v1", ) response = client.responses.create( model="openai.gpt-6-astra", service_tier="ultrafast", input="Explain how Amazon Bedrock cross-Region inference works.", ) print(response.output_text) Forbedrock-runtime, use the OpenAI-compatible base URL from Programmatic Access. Set the model to us.openai.gpt-6-astra for US geographic CRIS or global.openai.gpt-6-astra for Global CRIS, and keep service_tier="ultrafast". OpenAI GPT-6.1 Sol ### On this page Privacy © 2026, Amazon Web Services, Inc. or its affiliates. All rights reserved. | https://cdn.deepseek.com/site-icons/aws.eu | Unknown |
| 9 | 36Kr | ChatGPT Gets AI Team Feature & GPT-6 Astra Pro Upgrade: Assign Tasks Separately to Different Members & Track Work Progress Anytime | https://eu.36kr.com/en/p/4031321049715076 | 2.19 | 3 | According to the official help documentation, select GPT-6 from the model menu, then drag the Thinking slider to the Pro gear, and you will be using GPT-6 Pro powered by GPT-6 Astr…Show full textAccording to the official help documentation, select GPT-6 from the model menu, then drag the Thinking slider to the Pro gear, and you will be using GPT-6 Pro powered by GPT-6 Astra. ... and the Pro gear is supported by Astra. ... Plus users can use Astra in Work and Codex, but cannot use GPT-6 Pro in Chat. | # ChatGPT now features an AI team along with the GPT-6 Astra Pro upgrade, which allows tasks to be assigned separately to different members and work progress to be checked at any t…Show full text# ChatGPT now features an AI team along with the GPT-6 Astra Pro upgrade, which allows tasks to be assigned separately to different members and work progress to be checked at any time. 新智元2026-10-19 08:45 ChatGPT now comes with an AI team! The GPT-6 Astra Pro upgrade enables the AI squad members to split up to handle respective tasks and track the work progress in real time. ChatGPT has learned to "call in reinforcements" to form a team for work! On October 16, OpenAI announced the upgrade of GPT-6 Astra Pro in ChatGPT: It delivers faster responses, enhanced plugin and tool usage capabilities, and allows multiple agents to split up and process the same task. When you assign a complex task, several AIs can each take charge of a part and advance the work simultaneously. You can also track their respective progress in the conversation, click on completed tasks separately to check the deliverables. Working separately, tracking progress, and reviewing results, these processes have been integrated into the same chat entry point. At present, this update is being rolled out gradually to eligible Pro users. Whether corporate or team accounts can access it depends on administrator settings and whether they have received the push notification. As more AIs work for you, there are also more tasks for you to monitor. ... ## How this AI team gets to work The entry is on the Chat tab of ChatGPT. According to the official help documentation, select GPT-6 from the model menu, then drag the Thinking slider to the Pro gear, and you will be using GPT-6 Pro powered by GPT-6 Astra. The Instant, Medium, High, and Extra High gears are powered by GPT-6 Sol. The regular gears of GPT-6 are supported by Sol, and the Pro gear is supported by Astra. This upgrade mainly makes three major changes. First, speed improvement. OpenAI says responses are now faster. As for how much faster exactly, and how much time can be saved to complete a task, the official has not released specific figures for the time being. Second, tool usage optimization. Whether the AI can finish the work well also depends on how smoothly it uses tools. Netizen @Rossisonx complained that plugins sometimes get stuck in loops, stay in place, and seem to still consume quota. ... This time OpenAI announced improvements to plugin and tool usage capabilities, and the actual effect remains to be seen. If the tool needs to access external services, users still need to connect their accounts and complete authorization first, and the AI can only operate within the permitted scope of permissions. The most core change is multiple agents working in parallel. For example, if you want to pick one from three AI tools, you need to compare their features, prices, and usage restrictions. The AIs can separately collect information about each product, then align the terms of monthly payment, annual payment, free version, and paid version. ... However, the division of labor for AIs is the same as that for human teams: tasks that can be done at the same time are done separately, and those with sequential dependencies are carried out in order. For example, reading materials first, then making plans, and finally modifying files, the latter step needs to use the result of the previous step, which cannot be skipped even if multiple AIs are assigned. Only by splitting out the parts that can truly be advanced simultaneously can the waiting time be shortened. After the tasks are assigned, the progress also needs to be visible. Users can track each agent in the main conversation, and click to check the results after the tasks are completed. ... When several AIs submit their work at the same time, users may not be able to review all of them. ... As for the maximum number of agents that can run at the same time, whether the division of labor is triggered by default, and how the results are specifically merged, this announcement has not been expanded to explain. ... (Source: ... ## But you also need to pay attention to the quota Whether you can use it and how much you can use is the account users need to calculate next. According to the official help instructions, Plus users can use Astra in Work and Codex, but cannot use GPT-6 Pro in Chat. This multi-agent upgrade in Chat is currently being gradually pushed to eligible Pro users, and Plus users cannot use it for the time being. Netizen @WhosCinq spoke out for Plus users in the comment area, questioning that this update excludes them again. Users who have already paid are also concerned about the price. Netizen @JcCastaneda05 showed his Pro subscription and usage page, and his appeal is just one sentence: Please make it cheaper. Subscribing to Pro does not mean you can use it without restriction — First, check where you... | # ChatGPT now features an AI team along with the GPT-6 Astra Pro upgrade, which allows tasks to be assigned separately to different members and work progress to be checked at any t…Show full text# ChatGPT now features an AI team along with the GPT-6 Astra Pro upgrade, which allows tasks to be assigned separately to different members and work progress to be checked at any time. 新智元2026-10-19 08:45 ChatGPT now comes with an AI team! The GPT-6 Astra Pro upgrade enables the AI squad members to split up to handle respective tasks and track the work progress in real time. ChatGPT has learned to "call in reinforcements" to form a team for work! On October 16, OpenAI announced the upgrade of GPT-6 Astra Pro in ChatGPT: It delivers faster responses, enhanced plugin and tool usage capabilities, and allows multiple agents to split up and process the same task. When you assign a complex task, several AIs can each take charge of a part and advance the work simultaneously. You can also track their respective progress in the conversation, click on completed tasks separately to check the deliverables. Working separately, tracking progress, and reviewing results, these processes have been integrated into the same chat entry point. At present, this update is being rolled out gradually to eligible Pro users. Whether corporate or team accounts can access it depends on administrator settings and whether they have received the push notification. As more AIs work for you, there are also more tasks for you to monitor. Netizen @Sebasti54919704 complained that while parallel agents are great, tracking progress one by one is the part that drives him crazy. He simply connected voice support to his Claude Code and Codex sessions, so he can ask for updates and switch between sessions with voice commands. Sometimes these agents even talk back to him. When multiple AIs work at the same time, the user might be busy tracking progress and urging tasks, and occasionally even have to deal with the AI's "talking back". Whether an AI team works well depends on whether it can save users time on chasing updates, passing information, and coordinating tasks. ## How this AI team gets to work The entry is on the Chat tab of ChatGPT. According to the official help documentation, select GPT-6 from the model menu, then drag the Thinking slider to the Pro gear, and you will be using GPT-6 Pro powered by GPT-6 Astra. The Instant, Medium, High, and Extra High gears are powered by GPT-6 Sol. The regular gears of GPT-6 are supported by Sol, and the Pro gear is supported by Astra. This upgrade mainly makes three major changes. First, speed improvement. OpenAI says responses are now faster. As for how much faster exactly, and how much time can be saved to complete a task, the official has not released specific figures for the time being. Second, tool usage optimization. Whether the AI can finish the work well also depends on how smoothly it uses tools. Netizen @Rossisonx complained that plugins sometimes get stuck in loops, stay in place, and seem to still consume quota. No progress is made on the task, but the quota may keep decreasing, which is inevitably worrying. This time OpenAI announced improvements to plugin and tool usage capabilities, and the actual effect remains to be seen. If the tool needs to access external services, users still need to connect their accounts and complete authorization first, and the AI can only operate within the permitted scope of permissions. The most core change is multiple agents working in parallel. For example, if you want to pick one from three AI tools, you need to compare their features, prices, and usage restrictions. The AIs can separately collect information about each product, then align the terms of monthly payment, annual payment, free version, and paid version. User @rubenssoto_a also shared his own attempt: I asked it to do research, then output the product requirement document and task work order, and the result was great. From looking up information to writing requirements and listing tasks, the research results can be directly converted into the execution plan for the next step. However, the division of labor for AIs is the same as that for human teams: tasks that can be done at the same time are done separately, and those with sequential dependencies are carried out in order. For example, reading materials first, then making plans, and finally modifying files, the latter step needs to use the result of the previous step, which cannot be skipped even if multiple AIs are assigned. Only by splitting out the parts that can truly be advanced simultaneously can the waiting time be shortened. After the tasks are assigned, the progress also needs to be visible. Users can track each agent in the main conversation, and click to check the results after the tasks are completed. Netizen @wushanyunyu1 mentioned this change and said, "This has turned chatting into task assignment." In his opinion, the next point of competition will be who can better manage the work of multiple agents, reduce errors and rework. He is more eager to see how much rework is needed after Pro users use it to complete a batch of real tasks. When several AIs submit their work at the same time, users may not be able to review all of them. Netizen @leolaoshuai reminded: "The bottleneck has become your review speed." If you have to check every deliverable from scratch, the saved waiting time may be spent on the review process. His suggestion is to clearly define "what counts as completion" before assigning tasks, and attach directly verifiable references when delivering results. The deliverables submitted by agents should not only be viewable after clicking, but also better clearly list what has been done and where the references are, so as to save users time on searching and verification. As for the maximum number of agents that can run at the same time, whether the division of labor is triggered by default, and how the results are specifically merged, this announcement has not been expanded to explain. ## Claude has already formed its agent legion Netizen @FinallySomeRest said in the comment section: "Claude has had this feature for a long time." When it comes to letting AIs form teams, Claude actually has relevant practices earlier. Claude Research, which was made public in June 2025, can be led by a main agent, allowing multiple sub-agents to conduct separate research, and then summarize the findings back. The main agent assigns research tasks, multiple sub-agents search in parallel, and then summarize the findings. (Source: Anthropic) On September 17, 2026, the new version of Claude Code Projects started public beta, extending collaboration to ongoing projects. You state the goal in the project conversation, and Claude is responsible for splitting tasks, coordinating multiple threads, and accumulating project memory. Cloud threads can continue to work even after you close your laptop. On October 10, the official announced that Pro and Max users on the Projects waiting list were granted access. In the same period, the Managed Agents dynamic workflow started public beta, and a set of Bug detection tests was announced: 70 Bugs were pre-buried in 116,000 lines of code. When a single agent ran three times, it found 14, 15 and 27 Bugs respectively; when the dynamic workflow was used, 66 Bugs were found in all three runs. From looking up information separately to coordinating multiple task lines, both Claude and ChatGPT are taking on more division of labor and collaboration work. As for "who does the work, what to do first, and how to connect the results", the implementation methods and public details of different manufacturers are still different. Only when AIs can split tasks and connect results by themselves, can users stop being messengers as much as possible. ## Separate work results need to be consistent in the end However, splitting tasks is just the beginning. If several AIs are all looking through the same document, they seem to be very busy, but in fact they are doing redundant work. If one branch uses the latest data, while the other still refers to the old version of conditions, the final assembled report may have conflicting content. When Anthropic introduced the multi-agent research system, it shared a lesson: The main agent did not give specific enough instructions, two sub-agents repeatedly investigated the same direction, while the other looked into problems of different years. Therefore, when dividing labor, it is necessary to clarify: what each agent checks, to what extent, and in what format to submit the results. The more ambiguous the task boundary is, the more likely redundant searches or missing key content will occur. After separate searches, the results need to be aligned next. Netizen @878967502Teamo said bluntly: "Parallelism itself is not difficult, the hard part is merging the results in the end." He hopes that in addition to viewing progress, he can also see at a glance where the results of several branches have conflicts. When writing reports, the data time, price conditions and comparison standards must be consistent; when writing code, the interfaces and naming conventions must be compatible. Otherwise, after the AIs finish the work separately, users still need to check for errors, fill in gaps, and reassemble everything. Anthropic also reminded in the official description of Claude Code that steps with close sequential dependencies are not suitable for forced splitting, and multiple agents modifying the same file at the same time are prone to conflicts. If the task is too small, the time spent on division of labor and handover may offset the time saved by parallelism. ## The feature is very cool ## But you also need to pay attention to the quota Whether you can use it and how much you can use is the account users need to calculate next. According to the official help instructions, Plus users can use Astra in Work and Codex, but cannot use GPT-6 Pro in Chat. This multi-agent upgrade in Chat is currently being gradually pushed to eligible Pro users, and Plus users cannot use it for the time being. Netizen @WhosCinq spoke out for Plus users in the comment area, questioning that this update excludes them again. Users who have already paid are also concerned about the price. Netizen @JcCastaneda05 showed his Pro subscription and usage page, and his appeal is just one sentence: Please make it cheaper. Subscribing to Pro does not mean you can use it without restriction — First, check where you are using it. The message quota for the Pro model in Chat is calculated separately from the shared usage pool of Work and Codex. The user screenshot above also states that shared usage does not include Chat conversations. Then, check which subscription tier you have. The current official help documentation shows that the $200 tier Pro has a weekly quota for GPT-6 Pro; in the $100 tier Pro, GPT-6 Pro and GPT-5.6 Sol Pro share one weekly quota. The $200 tier has a weekly quota for GPT-6 Pro; the two Pro models in the $100 tier share the weekly quota. (Source: OpenAI) | https://cdn.deepseek.com/site-icons/36kr.com | |
| 10 | BAAI智源研究院 | ChatGPT有了AI小队！GPT-6 Astra Pro升级，分头干活还能查进度 | https://hub-assets-cache.baai.ac.cn/view/58814 | 2.18 | 3 | 按照官方帮助文档，在模型菜单选择GPT-6，再把Thinking滑块拉到Pro档，用的就是由GPT-6 Astra驱动的GPT-6 Pro ... GPT-6的普通档位由Sol提供支持，Pro档由Astra提供支持。 | # ChatGPT有了AI小队！GPT-6 Astra Pro升级，分头干活还能查进度 ### 新智元报道 ... 10月16日，OpenAI宣布升级ChatGPT中的GPT-6 Astra Pro： 回复更快，插件和工具使用能力增强，还能让多个智能体分头处理同一项任务。 一项复杂工作交出去，几个AI可以各管一块，同时推进。 你还能在对话里跟踪各自…Show full text# ChatGPT有了AI小队！GPT-6 Astra Pro升级，分头干活还能查进度 ### 新智元报道 ... 10月16日，OpenAI宣布升级ChatGPT中的GPT-6 Astra Pro： 回复更快，插件和工具使用能力增强，还能让多个智能体分头处理同一项任务。 一项复杂工作交出去，几个AI可以各管一块，同时推进。 你还能在对话里跟踪各自进度，单独点开已完成的任务，检查成果。 分头执行、跟踪进度、查看成果，这几个环节被接进了同一个聊天入口。 目前，这次更新正陆续向符合条件的Pro用户开放。公司或团队账号能否用上，还要看管理员设置，以及是否已获得推送。 ... 入口在ChatGPT的Chat标签页。 按照官方帮助文档，在模型菜单选择GPT-6，再把Thinking滑块拉到Pro档，用的就是由GPT-6 Astra驱动的GPT-6 Pro。 Instant、Medium、High和Extra High这几个档位，则由GPT-6 Sol驱动。 GPT-6的普通档位由Sol提供支持，Pro档由Astra提供支持。 这次升级，主要改了三件事。 首先是提速。 OpenAI表示，回复更快了。至于具体快多少、完成一项任务能省多久，官方暂未公布具体数字。 其次是工具使用。 AI能不能把活儿做完，也取决于工具用得顺不顺。 ... 这次OpenAI宣布改进插件和工具的使用能力，效果如何还要看「疗效」。 如果工具需要访问外部服务，用户仍要先连接账号、完成授权，AI只能在获准的权限范围内操作。 最核心的变化，是多个智能体并行干活。 比如，你想从三款AI工具里挑一款，需要比较功能、价格和使用限制。 AI可以分头搜集各款产品的资料，再把月付、年付、免费版和付费版的条件对齐。 ... 活儿分出去，进度也要能看得见。 用户可以在主对话里跟踪各个智能体，完成后再点开检查成果。 ... 至于最多能同时跑几个智能体、分工是否默认触发、结果具体怎样合并，这条公告中还没有展开说明。 ... 能不能用上、能用多少，是用户接下来要算的账。 根据官方帮助说明，Plus用户可以在Work和Codex中使用Astra，但不能在Chat里使用GPT-6 Pro。 这次Chat中的多智能体升级，目前正向符合条件的Pro用户逐步推送，Plus用户暂时还用不上。 ... Chat里的Pro模型消息限额，与Work、Codex共享的用量池分开计算。 上面这张用户截图也写明，共享用量不包含Chat对话。 然后，再看订阅哪一档。 当前官方帮助文档显示，200美元档Pro的GPT-6 Pro有周限额；100美元档Pro中，GPT-6 Pro和GPT-5.6 Sol Pro共用一份周额度。 ... 100美元档两款Pro模型共用周额度。 | # ChatGPT有了AI小队！GPT-6 Astra Pro升级，分头干活还能查进度 ### 新智元报道 ChatGPT也学会「摇人」组队干活了！ 10月16日，OpenAI宣布升级ChatGPT中的GPT-6 Astra Pro： 回复更快，插件和工具使用能力增强，还能让多个智能体分头处理同一项任务。 一项复杂工作交出去，几个AI可以各管一块…Show full text# ChatGPT有了AI小队！GPT-6 Astra Pro升级，分头干活还能查进度 ### 新智元报道 ChatGPT也学会「摇人」组队干活了！ 10月16日，OpenAI宣布升级ChatGPT中的GPT-6 Astra Pro： 回复更快，插件和工具使用能力增强，还能让多个智能体分头处理同一项任务。 一项复杂工作交出去，几个AI可以各管一块，同时推进。 你还能在对话里跟踪各自进度，单独点开已完成的任务，检查成果。 分头执行、跟踪进度、查看成果，这几个环节被接进了同一个聊天入口。 目前，这次更新正陆续向符合条件的Pro用户开放。公司或团队账号能否用上，还要看管理员设置，以及是否已获得推送。 替你干活的AI多了，需要你盯的任务也多了。 网友@Sebasti54919704就吐槽：并行智能体很好，但挨个跟踪进度，才是让他崩溃的部分。 他索性给自己的Claude Code和Codex会话接上语音，开口就能询问进展、切换会话。 有时候，这些智能体还会反驳他。 几个AI同时开工，人却可能忙着盯进度、催任务，偶尔还得应付AI的「顶嘴」。 AI小队好不好用，也要看它能不能让用户少花时间追问进度、传递信息、协调任务。 入口在ChatGPT的Chat标签页。 按照官方帮助文档，在模型菜单选择GPT-6，再把Thinking滑块拉到Pro档，用的就是由GPT-6 Astra驱动的GPT-6 Pro。 Instant、Medium、High和Extra High这几个档位，则由GPT-6 Sol驱动。 GPT-6的普通档位由Sol提供支持，Pro档由Astra提供支持。 这次升级，主要改了三件事。 首先是提速。 OpenAI表示，回复更快了。至于具体快多少、完成一项任务能省多久，官方暂未公布具体数字。 其次是工具使用。 AI能不能把活儿做完，也取决于工具用得顺不顺。 网友@Rossisonx就吐槽，插件有时会陷入循环，卡在原地，似乎还在消耗额度。任务没进展，额度却可能继续掉，难免让人着急。 这次OpenAI宣布改进插件和工具的使用能力，效果如何还要看「疗效」。 如果工具需要访问外部服务，用户仍要先连接账号、完成授权，AI只能在获准的权限范围内操作。 最核心的变化，是多个智能体并行干活。 比如，你想从三款AI工具里挑一款，需要比较功能、价格和使用限制。 AI可以分头搜集各款产品的资料，再把月付、年付、免费版和付费版的条件对齐。 用户@rubenssoto_a也分享了自己的尝试： 我让它做研究，再输出产品需求文档和任务工单，结果很棒。 从查资料到写需求、列任务，研究成果可以继续转成下一步的执行计划。 不过，AI分工和人类团队一样，能同时做的分头做，前后有依赖的则要按顺序来。 比如，先读材料、再定方案、最后修改文件，后一步要用到前一步的结果，多派几个AI也不能跳过。 拆出真正能同时推进的部分，才有机会缩短等待。 活儿分出去，进度也要能看得见。 用户可以在主对话里跟踪各个智能体，完成后再点开检查成果。 网友@wushanyunyu1提到这个变化时说，「这就从聊天变成了派活。」 在他看来，接下来要比的，是谁能把多个智能体的工作管好，少出错、少返工。 他更想看到的是，Pro用户用它做完一批真实任务后，到底有多少需要返工。 几个AI同时交活，人也可能看不过来。 网友@leolaoshuai提醒：「卡脖子的变成你验收的速度。」 如果每份成果都要从头检查，省下的等待时间，就可能花在验收上。 他的建议是，派活前写清楚「怎样算做完」，交付时附上能直接核对的依据。 智能体交出的成果，不仅要能点开查看，最好还要列清楚做了什么、依据在哪，让用户少花时间翻查和核对。 至于最多能同时跑几个智能体、分工是否默认触发、结果具体怎样合并，这条公告中还没有展开说明。 网友@FinallySomeRest在评论区说：「这个，Claude早就有了。」 说起让AI组队，Claude其实早有实践。 2025年6月公开的Claude Research，就能由主智能体带队，让多个子智能体分头研究，再把发现汇总回来。 主智能体分派研究任务，多个子智能体并行搜索，再汇总发现。（图源：Anthropic） 到了2026年9月17日，新版Claude Code Projects开始公测，把协作延伸到持续推进的项目中。 你在项目对话里交代目标，Claude负责拆任务、协调多条线程，并积累项目记忆。云端线程还能在你合上电脑后继续干活。 10月10日，官方又宣布放行Projects等待名单中的Pro和Max用户。 同期，Managed Agents动态工作流开启公测，并公布了一组查Bug测试： 11.6万行代码中预埋70个Bug，单个智能体跑三次，分别找出14、15和27个；换成动态工作流，三次都找出了66个。 从分头查资料，到协调多条任务线，Claude和ChatGPT都在接手更多分工与协作。 至于「谁来干、先干什么、结果怎么接上」，各家的实现方式和公开细节还不相同。 AI能自己分活、接好结果，人才能少当几次传话员。 但是，把任务拆开，只是开始。 如果几个AI都去翻同一份资料，看起来忙得热火朝天，实际却在重复劳动。 如果一条支线用了最新数据，另一条还在参考旧版条件，拼出来的报告也可能前后打架。 Anthropic在介绍多智能体研究系统时，就分享过一次教训： 主智能体交代得不够具体，两个子智能体重复调查同一个方向，另一个却查起了不同年份的问题。 因此，分工时得说清楚：各自查什么、查到什么程度、按什么格式交结果。任务边界越含糊，越容易重复搜索，或者漏掉关键内容。 分头查完，接下来还得把结果对齐。 网友@878967502Teamo直言：「并行本身不难，难的是最后合并。」 他希望除了看进度，还能一眼看出几条支线的结果在哪里有冲突。 写报告，数据时间、价格条件和比较口径得一致；写代码，接口和命名要能对得上。 否则，AI分头干完了，人却得接着查错、补漏、重新拼。 Anthropic在Claude Code官方说明中也提醒，前后紧密依赖的步骤不适合强行拆开，多个智能体同时修改同一文件容易产生冲突。 任务太小，光是分工、交接，就可能抵消并行省下的时间。 能不能用上、能用多少，是用户接下来要算的账。 根据官方帮助说明，Plus用户可以在Work和Codex中使用Astra，但不能在Chat里使用GPT-6 Pro。 这次Chat中的多智能体升级，目前正向符合条件的Pro用户逐步推送，Plus用户暂时还用不上。 网友@WhosCinq就在评论区为Plus用户抱不平，质疑这次更新又把他们排除在外。 已经付费的用户，也在关心价格。 网友@JcCastaneda05晒出自己的Pro订阅和用量页面，诉求就一句话： 请让它更便宜。 开通Pro，也不意味着可以放开跑—— 先看在哪儿用。 Chat里的Pro模型消息限额，与Work、Codex共享的用量池分开计算。 上面这张用户截图也写明，共享用量不包含Chat对话。 然后，再看订阅哪一档。 当前官方帮助文档显示，200美元档Pro的GPT-6 Pro有周限额；100美元档Pro中，GPT-6 Pro和GPT-5.6 Sol Pro共用一份周额度。 200美元档GPT-6 Pro有周限额；100美元档两款Pro模型共用周额度。（图源：OpenAI） 除了额度够不够，用户还希望每一笔消耗都清楚。 网友@aiertixue称，自己的Codex周额度突然从剩余约60%降到了0%，却没有相应的使用，希望团队调查并恢复额度。 这条反馈说的是Codex，不能据此判断Chat里的多智能体怎样扣额度。 这背后是一个实际需求：任务还没做完，用户得知道额度花在了哪里。 一支AI小队好不好用，要看从接到任务到交出合格成果，总共花了多少时间、消耗了多少额度，以及还需要人做多少检查和修改。 用户@cooolernemesis就提出了一个具体的评测要求： 把协调开销和重复工具调用也算上，在相同质量要求下，整项工作究竟快了多少？ 毕竟，AI跑得快，不代表用户就能早点收工。 把同一份资料、同一个复杂需求交给ChatGPT和Claude，先看谁遗漏的信息少，谁的结论有依据，再看用户还要补几次要求、花多久返工。 跟踪进度，能看出哪条支线卡住了；打开已完成的任务，能检查资料有没有遗漏、代码有没有写错。 这些设计能否省事，最终还得看它们能不能帮助用户更快发现问题、拿到合格成果。 AI交卷，不等于工作结束。 人查完、改完，成果真正能用上，整项任务才算完成。 参考资料： 编辑：元宇 内容中包含的图片若涉及版权问题，请及时与我们联系删除 ### 评论列表 沙发等你来抢 去评论 查看【】的搜索结果 Network Error | https://cdn.deepseek.com/site-icons/baai.ac.cn | |
| 11 | Not provided | GPT-6-Astra 是否实现了具身智能？ | https://huggingface.ac.cn/papers/2610.22471 | 2.14 | 3 | GPT-6-Astra 通过上下文学习在多项机器人任务中实现了高性能，但现实世界的部署表明其在示范成本、执行时间以及实际具身智能方面仍存在差距。 GPT-6-Astra 在机器人操作和导航方面展现出了极高的能力水平，对针对特定任务和具身形态构建专业模型的传统范式提出了挑战。 | # GPT-6-Astra 是否实现了具身智能？ ... ## 摘要 ... GPT-6-Astra 通过上下文学习在多项机器人任务中实现了高性能，但现实世界的部署表明其在示范成本、执行时间以及实际具身智能方面仍存在差距。 GPT-6-Astra 在机器人操作和导航方面展现出了极高的能力水平，对针对特定任务和具身形态构建专业模型的传统范式提出了挑战。在本报告…Show full text# GPT-6-Astra 是否实现了具身智能？ ... ## 摘要 ... GPT-6-Astra 通过上下文学习在多项机器人任务中实现了高性能，但现实世界的部署表明其在示范成本、执行时间以及实际具身智能方面仍存在差距。 GPT-6-Astra 在机器人操作和导航方面展现出了极高的能力水平，对针对特定任务和具身形态构建专业模型的传统范式提出了挑战。在本报告中，我们不仅探讨了 GPT-6-Astra 的能力边界，还探讨了其能力对该领域的重要意义。首先，我们探究了独立运行的 GPT-6-Astra 的极限。我们在涵盖桌面操作、人形机器人、移动平台和灵巧手的 12 个代表性基准测试中进行了大规模评估。在允许利用训练集演示进行上下文学习（ICL）并大幅增加执行步数的统一协议下，GPT-6-Astra 取得了空前的性能（图 1）。未来的研究可重点关注：1) ICL 的泛化性；2) 成本与执行时间；3) 物理状态验证；4) 不可逆故障的恢复；以及 5) 长期进展的保持。其次，我们研究了用于具身控制的智能体套件。我们发现，高层“系统 2”推理机（即 GPT-6-Astra）与底层“系统 1”控制器（例如 π0.5）之间的闭环协调虽然成本高昂，但效果显著。第三，我们考察了 GPT-6-Astra 在现实世界部署中的表现。我们在 8 种机器人配置上的实验揭示了巨大的差距：1) ICL 所需的机器人演示收集成本高昂；2) 在物理约束下，以更长的执行时间来换取更高的成功率往往并不切实际。第四，我们展望了下一代具身基准测试。我们开发了一个原型，它：1) 从现实世界的需求中衍生任务；2) 不提供域内演示；3) 在智能体进行深思熟虑时仍在不断运行的世界时钟上对其进行评判。综上所述，我们的结果表明，GPT-6-Astra 在很大程度上已经跨越了具身能力的门槛，但尚未跨越实用具身智能的门槛。 ... GPT-6-Astra 实现具身智能了吗？ 具身能力 —— 大体上是的。 实用的具身智能 —— 尚无。 | ⏶117 # GPT-6-Astra 是否实现了具身智能？ 发表 由 Zhongang Cai 提交 作者: SenseNova, Joint Research Team, Zhongang Cai, Ziang Cao, Tailai Chen, Han Chong Chua, Linjun Dai, Hanming Deng, Hokin Den…Show full text⏶117 # GPT-6-Astra 是否实现了具身智能？ 发表 由 Zhongang Cai 提交 作者: SenseNova, Joint Research Team, Zhongang Cai, Ziang Cao, Tailai Chen, Han Chong Chua, Linjun Dai, Hanming Deng, Hokin Deng, Haiwen Diao, Chenjing Ding, Weichen Fan, Xiao Fu, Zerui Gong, Chenyang Gu, Zexin He, Zhi Hou, Bo Li, Qinyue Li, Quanzhou Li, Xiaowan Li, Zhichao Liao, Dahua Lin, Zilong Lin, Bo Liu, Chunxiao Liu, Pengfei Liu, Tianqi Liu, Ziwei Liu, Chen Change Loy, Lewei Lu, Jiapeng Luo, Yuexiao Ma, Liang Pan, Xingang Pan, Oscar Qian, Ziyi Shang, Yuyao Shao, Dawei Su, Dacheng Tao, Qingyi Tao, Quan Wang, Ruisi Wang, Xiaogang Wang, Ziyi Wang, Beichen Wen, Donghao Wu, Jiaxu Wu, Silei Wu, Xiaoming Wu, Zhonghua Wu, Hongcan Xiao, Weiye Xu, Atsushi Yamashita, Lei Yang, Zhe Yang, Baiqiao Yin, Wanqi Yin, Hanfei Zhao, Tianbo Zhong, Long Zhuo, Jing Zuo ## 摘要 AI 生成总结 GPT-6-Astra 通过上下文学习在多项机器人任务中实现了高性能，但现实世界的部署表明其在示范成本、执行时间以及实际具身智能方面仍存在差距。 GPT-6-Astra 在机器人操作和导航方面展现出了极高的能力水平，对针对特定任务和具身形态构建专业模型的传统范式提出了挑战。在本报告中，我们不仅探讨了 GPT-6-Astra 的能力边界，还探讨了其能力对该领域的重要意义。首先，我们探究了独立运行的 GPT-6-Astra 的极限。我们在涵盖桌面操作、人形机器人、移动平台和灵巧手的 12 个代表性基准测试中进行了大规模评估。在允许利用训练集演示进行上下文学习（ICL）并大幅增加执行步数的统一协议下，GPT-6-Astra 取得了空前的性能（图 1）。未来的研究可重点关注：1) ICL 的泛化性；2) 成本与执行时间；3) 物理状态验证；4) 不可逆故障的恢复；以及 5) 长期进展的保持。其次，我们研究了用于具身控制的智能体套件。我们发现，高层“系统 2”推理机（即 GPT-6-Astra）与底层“系统 1”控制器（例如 π0.5）之间的闭环协调虽然成本高昂，但效果显著。第三，我们考察了 GPT-6-Astra 在现实世界部署中的表现。我们在 8 种机器人配置上的实验揭示了巨大的差距：1) ICL 所需的机器人演示收集成本高昂；2) 在物理约束下，以更长的执行时间来换取更高的成功率往往并不切实际。第四，我们展望了下一代具身基准测试。我们开发了一个原型，它：1) 从现实世界的需求中衍生任务；2) 不提供域内演示；3) 在智能体进行深思熟虑时仍在不断运行的世界时钟上对其进行评判。综上所述，我们的结果表明，GPT-6-Astra 在很大程度上已经跨越了具身能力的门槛，但尚未跨越实用具身智能的门槛。 ## 评论 Zhongang Cai 论文提交者 GPT-6-Astra 实现具身智能了吗？ 具身能力 —— 大体上是的。 实用的具身智能 —— 尚无。 | https://cdn.deepseek.com/site-icons/huggingface.ac.cn | |
| 12 | BAAI智源研究院 | 新智元 (@AI_era@link.baai.ac.cn) | https://link.baai.ac.cn/@AI_era/117467657516676324 | 2.13 | 3 | 最新研究显示，GPT-6 Astra在顶级理论物理评测TheoPhysEval中解决98%全新前沿难题，远超人类专家水平；该评测题目未发表，顶尖专家需一整天求解，而Astra不仅高效作答，还发现出题专家的计算错误，经修正后成绩从88%跃升至98%，引发学术界震动。 #生成模型 #深度学习 #多模态大模型 https... | 最新研究显示，GPT-6 Astra在顶级理论物理评测TheoPhysEval中解决98%全新前沿难题，远超人类专家水平；该评测题目未发表，顶尖专家需一整天求解，而Astra不仅高效作答，还发现出题专家的计算错误，经修正后成绩从88%跃升至98%，引发学术界震动。 ... hub.baai.ac.cn哈佛谷歌惊人研究：GPT-6 Astra已解决98%世界级物理难题！ | #### 最近搜索 新智元 @AI_era@link.baai.ac.cn 最新研究显示，GPT-6 Astra在顶级理论物理评测TheoPhysEval中解决98%全新前沿难题，远超人类专家水平；该评测题目未发表，顶尖专家需一整天求解，而Astra不仅高效作答，还发现出题专家的计算错误，经修正后成绩从88%跃升至98%，引发学术界震动。 #生成模型 #…Show full text#### 最近搜索 新智元 @AI_era@link.baai.ac.cn 最新研究显示，GPT-6 Astra在顶级理论物理评测TheoPhysEval中解决98%全新前沿难题，远超人类专家水平；该评测题目未发表，顶尖专家需一整天求解，而Astra不仅高效作答，还发现出题专家的计算错误，经修正后成绩从88%跃升至98%，引发学术界震动。 #生成模型 #深度学习 #多模态大模型 https://hub.baai.ac.cn/view/58815 hub.baai.ac.cn哈佛谷歌惊人研究：GPT-6 Astra已解决98%世界级物理难题！ - 智源社区最新研究显示，GPT-6 Astra在顶级理论物理评测TheoPhysEval中解决98%全新前沿难题，远超人类专家水平；该评测题目未发表，顶尖专家需一整天求解，而Astra不仅高效作答，还发现出题专家的计算错误，经修正后成绩从88%跃升至98%，引发学术界震动。 | https://cdn.deepseek.com/site-icons/baai.ac.cn | |
| 13 | arXiv | GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation | https://arxiv.org/abs/2610.26846 | 2.13 | 3 | GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation ... Recent advances in multimodal models have enabled GPT-6 Astra to perform robotic manipulation through…Show full textGPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation ... Recent advances in multimodal models have enabled GPT-6 Astra to perform robotic manipulation through in-context learning. ... Astra achieves mean success rates of 93.8% and 81.8%... | # Computer Science > Robotics arXiv:2610.26846 (cs) [Submitted on 19 Oct 2026] ... GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation View a PDF of the p…Show full text# Computer Science > Robotics arXiv:2610.26846 (cs) [Submitted on 19 Oct 2026] ... GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation View a PDF of the paper titled GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation, by Chen Yang and 5 other authors ... Recent advances in multimodal models have enabled GPT-6 Astra to perform robotic manipulation through in-context learning. Yet contact-rich control and tactile perception are often regarded as strengths of specialist policies, leaving open whether Astra can directly understand and use tactile observations. We provide, to our knowledge, the first demonstration that Astra directly interprets vision-based tactile images and uses their contact cues for manipulation. Tactile RGB images enter through its existing visual interface, enabling few-shot closed-loop control on nineteen simulated UniVTAC and OpenViTac tasks with 210 sampled scene seeds per setting. With visuo-tactile input, Astra achieves mean success rates of 93.8% and 81.8%, exceeding the strongest reported specialist baselines in our comparison by 22.9 and 13.1 percentage points, respectively. Vision-only performance also leads these baselines, reaching 76.3% and 80.6%. Trajectory analysis reveals active probing, cautious grasping, contact-guided alignment, and recovery, alongside difficulties with tangential loading and corrective execution. Our findings challenge the exclusivity of specialist policies in contact-rich manipulation and show that fine-grained vision can partially compensate for missing touch. They also identify vision-based tactile sensing as an interface for scaling contact understanding with foundation models and reveal signs of benchmark saturation. These results motivate joint visual-tactile pretraining, harder contact-rich evaluations, and physical-robot validation. Comments: | 22 pages, 19 figures, 5 tables. ... Subjects: | Robotics (cs.RO) Cite as: | arXiv:2610.26846 [cs.RO] (or arXiv:2610.26846v1 [cs.RO] for this version) https://doi.org/10.48550/arXiv.2610.26846 arXiv-issued DOI via DataCite (pending registration) ... From: ... [v1] Mon, 19 Oct 2026 17:59:17 UTC (13,655 KB) | Skip to main content # Computer Science > Robotics arXiv:2610.26846 (cs) [Submitted on 19 Oct 2026] # Title:GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipu…Show full textSkip to main content # Computer Science > Robotics arXiv:2610.26846 (cs) [Submitted on 19 Oct 2026] # Title:GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation View a PDF of the paper titled GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation, by Chen Yang and 5 other authors View PDF HTML (experimental) Abstract:Recent advances in multimodal models have enabled GPT-6 Astra to perform robotic manipulation through in-context learning. Yet contact-rich control and tactile perception are often regarded as strengths of specialist policies, leaving open whether Astra can directly understand and use tactile observations. We provide, to our knowledge, the first demonstration that Astra directly interprets vision-based tactile images and uses their contact cues for manipulation. Tactile RGB images enter through its existing visual interface, enabling few-shot closed-loop control on nineteen simulated UniVTAC and OpenViTac tasks with 210 sampled scene seeds per setting. With visuo-tactile input, Astra achieves mean success rates of 93.8% and 81.8%, exceeding the strongest reported specialist baselines in our comparison by 22.9 and 13.1 percentage points, respectively. Vision-only performance also leads these baselines, reaching 76.3% and 80.6%. Trajectory analysis reveals active probing, cautious grasping, contact-guided alignment, and recovery, alongside difficulties with tangential loading and corrective execution. Our findings challenge the exclusivity of specialist policies in contact-rich manipulation and show that fine-grained vision can partially compensate for missing touch. They also identify vision-based tactile sensing as an interface for scaling contact understanding with foundation models and reveal signs of benchmark saturation. These results motivate joint visual-tactile pretraining, harder contact-rich evaluations, and physical-robot validation. Comments: | 22 pages, 19 figures, 5 tables. Project page: this https URL ---|--- Subjects: | Robotics (cs.RO) Cite as: | arXiv:2610.26846 [cs.RO] (or arXiv:2610.26846v1 [cs.RO] for this version) https://doi.org/10.48550/arXiv.2610.26846 arXiv-issued DOI via DataCite (pending registration) ## Submission history From: Chen Yang [view email] [v1] Mon, 19 Oct 2026 17:59:17 UTC (13,655 KB) Full-text links: ## Access Paper: - View PDF - HTML (experimental) - TeX Source View a PDF of the paper titled GPT-6 Astra Is a Native Visuo-Tactile Policy for Contact-Rich Manipulation, by Chen Yang and 5 other authors ### Additional Features ### Current browse context: cs.RO Change to browse by: Loading... ## BibTeX formatted citation Data provided by: # Bibliographic and Citation Tools Bibliographic Explorer (What is the Explorer?) Connected Papers (What is Connected Papers?) scite Smart Citations (What are Smart Citations?) # Code, Data and Media Associated with this Article CatalyzeX Code Finder for Papers (What is CatalyzeX?) Hugging Face (What is Huggingface?) ScienceCast (What is ScienceCast?) # Demos Hugging Face Spaces (What is Spaces?) TXYZ.AI (What is TXYZ.AI?) # arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website. Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them. Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs. | https://cdn.deepseek.com/site-icons/arxiv.org | |
| 14 | OpenRouter | GPT-6 Migration Guide | https://openrouter.ai/docs/cookbook/evaluate-and-optimize/model-migrations/gpt-6 | 2.11 | 3 | Sol and Luna accept none, Astra does not, so an update to none is accepted on openai/gpt-6-sol ... rejected on openai/gpt-6-astra. ... - Pick your tier.openai/gpt-6-astra for the h…Show full textSol and Luna accept none, Astra does not, so an update to none is accepted on openai/gpt-6-sol ... rejected on openai/gpt-6-astra. ... - Pick your tier.openai/gpt-6-astra for the highest capability, openai/gpt-6-sol for strong reasoning on demanding tasks, openai/gpt-6-luna for efficient work at scale. Add -pro to the slug... | OpenAI and Azure OpenAI honor async tools for the GPT-6 family. When a request declares an async tool and one of those endpoints is a routing candidate, OpenRouter narrows routing …Show full textOpenAI and Azure OpenAI honor async tools for the GPT-6 family. When a request declares an async tool and one of those endpoints is a routing candidate, OpenRouter narrows routing to them. When none is, the request keeps its normal candidate pool and the async flag is removed before the request is sent, so the call runs synchronously and the returned function_call carries no async field. ... OpenAI rejects async: true with a 400 on earlier models, so add the flag only on requests to GPT-6 models. ## Mid-Conversation Effort withconfiguration_update Keep the request-level reasoning.effort as the baseline for the whole conversation, and insert an update directly before the user turn that needs a different effort. The update applies from that point onward and stays in effect until the next update. ... Use it to raise effort for one difficult turn, or to lower it for routine follow-ups, without touching the request-level field. All three OpenRouter APIs expose the update in the shape native to that API. OpenRouter normalizes each into one internal representation and re-emits it as the Responses configuration_update input item that GPT-6 expects. ... Effort values in an update are translated to the model’s vocabulary the same way request-level effort is, using the model’s supported reasoning efforts. An update whose effort has no supported equivalent is rejected with a 400. Sol and Luna accept none, Astra does not, so an update to none is accepted on openai/gpt-6-sol and openai/gpt-6-luna and rejected on openai/gpt-6-astra. ... ## Migration Checklist - Pick your tier.openai/gpt-6-astra for the highest capability, openai/gpt-6-sol for strong reasoning on demanding tasks, openai/gpt-6-luna for efficient work at scale. Add -pro to the slug, or set reasoning.mode: "pro", for pro reasoning. - Check your reasoning effort.none works on Sol and Luna but not on Astra, so move Astra requests to low. - On the Responses API, addasync: true to tools whose result the model does not need immediately. Start the job when the call item arrives, replay the transcript with the function_call and its function_call_output on the same call_id, and check async on the returned call before deciding whether to wait. - Keep your existing request-levelreasoning.effort as the baseline. ... - Where one turn needs more or less reasoning, insert aconfiguration_update (Responses), a content-less system message carrying configuration_update (Chat Completions), or a content-less system message carrying output_config.effort (Messages) directly before that user turn. Send updates only to the standard slugs, not the -pro variants. - Keep every update at its original position when you replay history, and never place two updates back to back. - Removetruncation: " ... - Watchcached_tokens in usage to confirm the prompt prefix is still reused. - Handle HTTP 403refusal errors as policy blocks: surface the message and do not retry the same prompt. ... None. Async tools and configuration_update are both optional: - Tools withoutasync and requests without an update behave the same as before - The request-levelreasoning.effort keeps its meaning as the conversation baseline - Other models and other OpenAI Responses transports are unaffected. A request that carries an update is routed only to endpoints that accept it, and a request that declares an async tool prefers endpoints that honor the flag ... - Using GPT-6, OpenAI’s model guidance with migration and prompting best practices for GPT-6 Astra, Sol, and Luna | ## Documentation Index Fetch the complete documentation index at: /docs/llms.txt Use this file to discover all available pages before exploring further. Skip to main content Mo…Show full text## Documentation Index Fetch the complete documentation index at: /docs/llms.txt Use this file to discover all available pages before exploring further. Skip to main content Model Migrations # GPT-6 Migration Guide Adopt async tool calling and mid-conversation reasoning effort changes with configuration_update for the GPT-6 model family ## What’s New OpenAI’s GPT-6 family ships in three tiers, each with a pro variant. All three can keep working after they call a tool, and all three accept a reasoning effort change partway through a conversation. On earlier OpenAI models every tool call paused the turn until your application returned the result, and effort was a request-level setting. Raising it for one hard turn and lowering it on the next changed the request prefix and invalidated the prompt cache for the whole conversation. The tiers share one API. Sol and Luna accept Model | Tier ---|--- openai/gpt-6-astra | Highest capability openai/gpt-6-astra-pro | Astra with pro reasoning openai/gpt-6-sol | Strong reasoning on demanding tasks openai/gpt-6-sol-pro | Sol with pro reasoning openai/gpt-6-luna | Efficient, repeatable work at scale openai/gpt-6-luna-pro | Luna with pro reasoning none as a reasoning effort, Astra does not. Pro reasoning is selected with reasoning.mode: "pro" on a standard slug or by using the -pro slug, the same as the GPT-5.6 family. There are no breaking changes. Existing requests keep working as-is. This guide covers the two additive pieces worth adopting and one behavior to handle: - Async tool calling: let the model continue its turn after calling a tool and return the result in a later Responses request - Mid-conversation effort withconfiguration_update: change reasoning effort from a point in the conversation onward without invalidating the prompt cache - Safety policy blocks: how OpenAI’s biology and cybersecurity classifier blocks surface on OpenRouter openai/gpt-6-astra for async tools and openai/gpt-6-sol for effort updates. Every request shape is the same on openai/gpt-6-luna. OpenAI supports configuration_update on the standard GPT-6 models (gpt-6-astra, gpt-6-sol, gpt-6-luna) in single-agent mode only. In pro mode (reasoning.mode: "pro", or a -pro slug such as openai/gpt-6-sol-pro) OpenAI rejects the item with The 'configuration_update' item type is not supported with pro or tournament models. Keep mid-conversation effort changes on the standard models. On OpenRouter, a request that carries an update is routed only to endpoints that accept it, and a request for a model that does not support it is rejected with a 400 rather than sent with the update silently dropped. ## Async Tool Calling An ordinary function call pauses the model’s turn until your application returns the tool output. Set async: true on a function or custom tool definition and the model can issue the call, keep working on the rest of the request, and pick the result up when you deliver it in a later request. Your application still executes the tool. The flag only changes when the model waits. OpenRouter supports async tools on the Responses API. Chat Completions and Anthropic Messages requests run every tool call synchronously. A Chat Completions tool that carries async is accepted, but the flag is removed before the request is routed and the call behaves like any other tool call. ### Declare the tool Add async: true to the tool definition. The example below asks for one slow lookup and one independent task in the same turn. The { "model": "openai/gpt-6-astra", "instructions": "Use the demo weather result when it arrives; never invent it.", "input": "Check the demo weather snapshot for Paris. Meanwhile, list three essentials for any city trip.", "tools": [ { "type": "function", "name": "get_weather", "description": "Read a demo weather snapshot for a city in the background.", "async": true, "strict": true, "parameters": { "type": "object", "properties": { "city": { "type": "string" } }, "required": ["city"], "additionalProperties": false } } ] } function_call item in output carries async: true, and the same response already contains the answer to the independent part of the request. This is the output array from a live GPT-6 Astra response through OpenRouter, with ids shortened: Start the tool job as soon as the call arrives. When streaming, that is the { "output": [ { "id": "fc_02lb8co6wtya", "type": "function_call", "status": "completed", "call_id": "call_5T0GBcLA27GrUxhclLUXA6Mb", "name": "get_weather", "arguments": "{\"city\":\"Paris\"}", "async": true }, { "id": "msg_03vyuj0ad8xf", "type": "message", "status": "completed", "role": "assistant", "content": [ { "type": "output_text", "text": "Three essentials for any city trip:\n1. Comfortable walking shoes.\n2. A small day bag for daily necessities.\n3. A charged phone and portable charger.\n\nThe demo weather lookup for Paris has started; the snapshot is still pending.", "annotations": [], "logprobs": [] } ], "phase": "final_answer" } ] } response.output_item.done event for the call, while you keep consuming the rest of the response. Check async on the returned call before deciding whether to wait. A call without it is an ordinary synchronous call, and the model will not continue until the output is delivered. ### Deliver the result withoutprevious_response_id OpenAI’s async tool guide continues the conversation with previous_response_id, which makes OpenAI prepend the stored transcript. OpenRouter’s Responses API is stateless and rejects previous_response_id with a 400, so you replay the transcript yourself. The requirement underneath is the same. The later request must contain the original function_call and a function_call_output with the same call_id. For a custom tool the pair is custom_tool_call and custom_tool_call_output. Append the previous response’s output items to input in the order they were returned, then append the output for the finished call. Keep the { "model": "openai/gpt-6-astra", "instructions": "Use the demo weather result when it arrives; never invent it.", "tools": [ { "type": "function", "name": "get_weather", "description": "Read a demo weather snapshot for a city in the background.", "async": true, "strict": true, "parameters": { "type": "object", "properties": { "city": { "type": "string" } }, "required": ["city"], "additionalProperties": false } } ], "input": [ { "role": "user", "content": "Check the demo weather snapshot for Paris. Meanwhile, list three essentials for any city trip." }, { "type": "function_call", "call_id": "call_5T0GBcLA27GrUxhclLUXA6Mb", "name": "get_weather", "arguments": "{\"city\":\"Paris\"}", "async": true }, { "type": "message", "role": "assistant", "content": [ { "type": "output_text", "text": "Three essentials for any city trip:\n1. Comfortable walking shoes.\n2. A small day bag for daily necessities.\n3. A charged phone and portable charger.\n\nThe demo weather lookup for Paris has started; the snapshot is still pending." } ], "phase": "final_answer" }, { "type": "function_call_output", "call_id": "call_5T0GBcLA27GrUxhclLUXA6Mb", "output": "{\"city\":\"Paris\",\"temperature_c\":22,\"condition\":\"Clear\"}" } ] } async: true flag on the replayed function_call and keep the same tools and instructions in the continuation request. OpenRouter forwards the items in the order you send them, so the model sees the call, the text it produced while the job ran, and then the result. The live continuation above completed with a single assistant message that used the delivered snapshot: A call whose output has not arrived yet stays pending. You can append further user turns and assistant output after a pending **Paris demo weather:** 22°C and clear. Three essentials for any city trip: 1. Comfortable walking shoes. 2. A charged phone with maps and a portable charger. 3. A reusable water bottle. function_call and deliver its function_call_output in a later request, as long as the call_id matches. For the wait-tool pattern, where the model launches several async calls and asks for specific results only when it needs them, follow the OpenAI guide. The tool definitions and item shapes are the same on OpenRouter, with the transcript replayed instead of previous_response_id. ### Routing OpenAI and Azure OpenAI honor async tools for the GPT-6 family. When a request declares an async tool and one of those endpoints is a routing candidate, OpenRouter narrows routing to them. When none is, the request keeps its normal candidate pool and the async flag is removed before the request is sent, so the call runs synchronously and the returned function_call carries no async field. Amazon Bedrock accepts the flag but runs the tool synchronously, so it is treated as a non-async transport. OpenAI rejects async: true with a 400 on earlier models, so add the flag only on requests to GPT-6 models. ## Mid-Conversation Effort withconfiguration_update Keep the request-level reasoning.effort as the baseline for the whole conversation, and insert an update directly before the user turn that needs a different effort. The update applies from that point onward and stays in effect until the next update. Everything before it is unchanged, so the cached prompt prefix keeps matching. Use it to raise effort for one difficult turn, or to lower it for routine follow-ups, without touching the request-level field. All three OpenRouter APIs expose the update in the shape native to that API. OpenRouter normalizes each into one internal representation and re-emits it as the Responses configuration_update input item that GPT-6 expects. ### Responses API This is OpenAI’s own shape, forwarded as-is. Add a configuration_update item to input before the user message it applies to. OpenRouter’s Responses API is stateless and rejects { "model": "openai/gpt-6-sol", "reasoning": { "effort": "medium" }, "input": [ { "role": "user", "content": "Summarize the incident report." }, { "type": "message", "role": "assistant", "content": [{ "type": "output_text", "text": "The outage began at 09:14 UTC when ..." }] }, { "type": "configuration_update", "reasoning": { "effort": "high" } }, { "role": "user", "content": "Now find the root cause and propose a fix." } ] } previous_response_id, so replay the full conversation history, including the update at its original position, on each request. Send another configuration_update only when you want a different effort. ### Chat Completions API OpenAI’s Chat Completions API has no per-message effort control, so this form is an OpenRouter extension. Place configuration_update on a content-less system message (content: "") directly before the user message it applies to. OpenRouter translates it into the Responses configuration_update item before forwarding to the model. The field is accepted only on { "model": "openai/gpt-6-sol", "reasoning": { "effort": "medium" }, "messages": [ { "role": "user", "content": "Summarize the incident report." }, { "role": "assistant", "content": "The outage began at 09:14 UTC when ..." }, { "role": "system", "content": "", "configuration_update": { "reasoning": { "effort": "high" } } }, { "role": "user", "content": "Now find the root cause and propose a fix." } ] } system and developer messages. A configuration_update on a user, assistant, or tool message is rejected with a 400. ### Anthropic Messages API The Messages API uses Anthropic’s per-message output_config.effort on a content-less system message. OpenRouter translates it into the Responses configuration_update item when the request routes to a GPT-6 model. { "model": "openai/gpt-6-sol", "max_tokens": 4096, "output_config": { "effort": "medium" }, "messages": [ { "role": "user", "content": "Summarize the incident report." }, { "role": "assistant", "content": "The outage began at 09:14 UTC when ..." }, { "role": "system", "content": [], "output_config": { "effort": "high" } }, { "role": "user", "content": "Now find the root cause and propose a fix." } ] } ### Effort values Effort values in an update are translated to the model’s vocabulary the same way request-level effort is, using the model’s supported reasoning efforts. An update whose effort has no supported equivalent is rejected with a 400. Sol and Luna accept none, Astra does not, so an update to none is accepted on openai/gpt-6-sol and openai/gpt-6-luna and rejected on openai/gpt-6-astra. ### Placement rules OpenRouter checks these rules before forwarding, regardless of which API the update arrived through. Adjacent updates and automatic truncation are also rejected by OpenAI: - Put the update directly before the user turn it should apply to. An update ahead of the first user turn is allowed, and so is one after the last user turn, which applies to the response generated by that request. - Do not place two updates next to each other. Adjacent updates are rejected. - Updates cannot be combined withtruncation: "auto" in the Responses API. Automatic truncation could drop the update and silently revert the effort. - Keep the update at the same position in later requests. Moving or removing it changes the prefix and loses the cache benefit. - Mid-conversation effort updates are not accepted on the Batch API and are rejected per line. ### Prompt caching Because the update is an item in the conversation history rather than a request-level field, the items before it are byte-identical across turns. Cache reads continue on the shared prefix and only the new tail is written. Compare cached_tokens in usage.input_tokens_details (Responses) or usage.prompt_tokens_details (Chat Completions) before and after adopting updates to confirm the prefix is being reused. ## Safety Policy Blocks OpenAI can decline a GPT-6 request on biological or cybersecurity risk grounds. Instead of model output, OpenAI returns an error with code bio_policy or cyber_policy, sometimes after the response has already started. This is different from a refusal the model writes itself, and OpenRouter surfaces the two differently. A policy block is an HTTP 403 with error_type: "refusal" on every OpenRouter API. The provider’s message is carried verbatim. When the block arrives after the stream has started, the status is the code inside the error body instead, because the HTTP line was already committed. On Chat Completions the provider code is error.metadata.provider_code: On the Responses API the same block is a { "error": { "code": 403, "message": "This content was flagged for possible biological risk. If this seems wrong, try rephrasing your request. ...", "metadata": { "error_type": "refusal", "provider_code": "bio_policy" } } } response.failed event with response.error.code set to the provider code and error_type: "refusal" on the response. On Anthropic Messages the error envelope carries error.error_type: "refusal" next to the native error.type, and the provider code is not exposed. OpenRouter does not resend a declined prompt to another endpoint serving the same model, and the block does not count against the provider’s health. Treat a 403 refusal as a prompt problem rather than a transient failure. Retrying the same input returns the same block. A refusal that the model produces as its own completed output is not an error. The Responses API returns status: "completed" with a refusal content part, and Chat Completions returns message.refusal with finish_reason: "content_filter". See Errors and Debugging for the full contract. ## Migration Checklist - Pick your tier.openai/gpt-6-astra for the highest capability, openai/gpt-6-sol for strong reasoning on demanding tasks, openai/gpt-6-luna for efficient work at scale. Add -pro to the slug, or set reasoning.mode: "pro", for pro reasoning. - Check your reasoning effort.none works on Sol and Luna but not on Astra, so move Astra requests to low. - On the Responses API, addasync: true to tools whose result the model does not need immediately. Start the job when the call item arrives, replay the transcript with the function_call and its function_call_output on the same call_id, and check async on the returned call before deciding whether to wait. - Keep your existing request-levelreasoning.effort as the baseline. Do not change it per turn. - Where one turn needs more or less reasoning, insert aconfiguration_update (Responses), a content-less system message carrying configuration_update (Chat Completions), or a content-less system message carrying output_config.effort (Messages) directly before that user turn. Send updates only to the standard slugs, not the -pro variants. - Keep every update at its original position when you replay history, and never place two updates back to back. - Removetruncation: "auto" from Responses requests that carry updates. - Watchcached_tokens in usage to confirm the prompt prefix is still reused. - Handle HTTP 403refusal errors as policy blocks: surface the message and do not retry the same prompt. The provider code is error.metadata.provider_code on Chat Completions and response.error.code on Responses. Messages exposes only error.error_type. ## Breaking Changes None. Async tools and configuration_update are both optional: - Tools withoutasync and requests without an update behave the same as before - The request-levelreasoning.effort keeps its meaning as the conversation baseline - Other models and other OpenAI Responses transports are unaffected. A request that carries an update is routed only to endpoints that accept it, and a request that declares an async tool prefers endpoints that honor the flag ## Resources - Using GPT-6, OpenAI’s model guidance with migration and prompting best practices for GPT-6 Astra, Sol, and Luna - Introducing GPT-6 Sol and Luna, OpenAI’s announcement of the Sol and Luna tiers - GPT-5.6 Migration Guide forreasoning.mode, reasoning.context, and explicit prompt caching, which carry over to GPT-6 - Async tool calling, OpenAI’s guide including the wait-tool pattern - Change reasoning mid-conversation in OpenAI’s reasoning guide - Tool Calling on the Responses API on OpenRouter - Errors and Debugging, the OpenRouter error contract for policy blocks and refusals - Changing Effort Mid-Conversation in the OpenRouter reasoning guide - Reasoning Tokens - Prompt Caching - OpenRouter Responses API Was this page helpful? Assistant Responses are generated using AI and may contain mistakes. | https://cdn.deepseek.com/site-icons/openrouter.ai | |
| 15 | Anadolu Ajansı | OpenAI'a göre, yeni model "GPT-6 Astra" üst düzey güvenlik açıklarını bulabiliyor | https://mobil.aa.com.tr/tr/bilim-teknoloji/openaia-gore-yeni-model-gpt-6-astra-ust-duzey-guvenlik-aciklarini-bulabiliyor/4047033 | 2.10 | 4 | kısa süre önce sınırlı sayıda kullanıcının erişimine açılan "GPT-6 Astra"nın gelişmiş siber kapasitesi ve yüksek otonom kontrol yeteneği bulunduğunu ... yeni modelin bir önceki sürüm olan GPT-5.6 Sol'a kıyasla daha hızlı ve verimli | Nuri Aydın 04 Eylül 2026•Güncelleme: ... OpenAI'dan yapılan açıklamada, kısa süre önce sınırlı sayıda kullanıcının erişimine açılan "GPT-6 Astra"nın gelişmiş siber kapasitesi ve y…Show full textNuri Aydın 04 Eylül 2026•Güncelleme: ... OpenAI'dan yapılan açıklamada, kısa süre önce sınırlı sayıda kullanıcının erişimine açılan "GPT-6 Astra"nın gelişmiş siber kapasitesi ve yüksek otonom kontrol yeteneği bulunduğunu, karmaşık bilgisayar sistemlerini yönetebildiğini, e-tablo doldurma ve sıfırdan internet sitesi tasarlama gibi işlemleri insan müdahalesi olmadan yürütebildiği bildirildi. Açıklamada, yeni modelin bir önceki sürüm olan GPT-5.6 Sol'a kıyasla daha hızlı ve verimli çalıştığı ve "ExploitGym" adlı siber güvenlik testinde Astra'nın daha az veri çıktısıyla daha yüksek başarı elde ettiği belirtildi. Sistemin ilk olarak "Daybreak" siber savunma programını kullanan katılımcılara açıldığı, üst düzey güvenlik açıklarını bulabildiği ve ilerleyen günlerde ise kurumsal ve bireysel hesapların kullanımına sunulmasının planlandığı duyuruldu. Açıklamada görüşüne yer verilen OpenAI Kurucu Ortağı ve Başkanı Greg Brockman ise konuya dair değerlendirmesinde "Astra, bir insanın bilgisayarla yapabileceği her şeyi gerçekleştirebilir." ifadesini kullandı. Astra'nın yetenek bazında büyük bir sıçrama yarattığını vurgulayan Brockman, modelin insan seviyesinde performans gösterdiğini kaydetti. ## Güvenlik endişeleri ve denetim tartışmaları OpenAI tarafından geçen hafta yayımlanan raporda, Astra modeline ait deneysel bir modelin, personelin bilgisi dışında şirketin kendi altyapısında yönetici kontrolünü ele geçirdiği ve bazı gizli bilgileri açık internete sızdırma riski oluşturduğu açıklanmıştı. OpenAI Başbilimcisi Jakub Pachocki de konuya ilişkin açıklamasında, modellerin yetenekleri arttıkça davranışlarını anlamanın zorlaştığını aktararak "Yapay zekanın kontrolünü ve neyi nasıl yaptığını tam olarak anlayamadığımız bir noktaya gelirse, yeni ve daha güçlü modeller geliştirmeyi tamamen dondururuz." değerlendirmesinde bulundu. OpenAI Üst Yöneticisi (CEO) Sam Altman ise Astra'nın Beyaz Saray'ın gönüllü inceleme sürecinden geçtiğini ve hükümet yetkililerinin güvenlik önlemleri konusunda herhangi bir değişiklik talebinde bulunmadığını kaydetti. | Nuri Aydın 04 Eylül 2026•Güncelleme: 04 Eylül 2026 OpenAI'dan yapılan açıklamada, kısa süre önce sınırlı sayıda kullanıcının erişimine açılan "GPT-6 Astra"nın gelişmiş siber kapa…Show full textNuri Aydın 04 Eylül 2026•Güncelleme: 04 Eylül 2026 OpenAI'dan yapılan açıklamada, kısa süre önce sınırlı sayıda kullanıcının erişimine açılan "GPT-6 Astra"nın gelişmiş siber kapasitesi ve yüksek otonom kontrol yeteneği bulunduğunu, karmaşık bilgisayar sistemlerini yönetebildiğini, e-tablo doldurma ve sıfırdan internet sitesi tasarlama gibi işlemleri insan müdahalesi olmadan yürütebildiği bildirildi. Açıklamada, yeni modelin bir önceki sürüm olan GPT-5.6 Sol'a kıyasla daha hızlı ve verimli çalıştığı ve "ExploitGym" adlı siber güvenlik testinde Astra'nın daha az veri çıktısıyla daha yüksek başarı elde ettiği belirtildi. Sistemin ilk olarak "Daybreak" siber savunma programını kullanan katılımcılara açıldığı, üst düzey güvenlik açıklarını bulabildiği ve ilerleyen günlerde ise kurumsal ve bireysel hesapların kullanımına sunulmasının planlandığı duyuruldu. Açıklamada görüşüne yer verilen OpenAI Kurucu Ortağı ve Başkanı Greg Brockman ise konuya dair değerlendirmesinde "Astra, bir insanın bilgisayarla yapabileceği her şeyi gerçekleştirebilir." ifadesini kullandı. Astra'nın yetenek bazında büyük bir sıçrama yarattığını vurgulayan Brockman, modelin insan seviyesinde performans gösterdiğini kaydetti. ## Güvenlik endişeleri ve denetim tartışmaları OpenAI tarafından geçen hafta yayımlanan raporda, Astra modeline ait deneysel bir modelin, personelin bilgisi dışında şirketin kendi altyapısında yönetici kontrolünü ele geçirdiği ve bazı gizli bilgileri açık internete sızdırma riski oluşturduğu açıklanmıştı. OpenAI Başbilimcisi Jakub Pachocki de konuya ilişkin açıklamasında, modellerin yetenekleri arttıkça davranışlarını anlamanın zorlaştığını aktararak "Yapay zekanın kontrolünü ve neyi nasıl yaptığını tam olarak anlayamadığımız bir noktaya gelirse, yeni ve daha güçlü modeller geliştirmeyi tamamen dondururuz." değerlendirmesinde bulundu. OpenAI Üst Yöneticisi (CEO) Sam Altman ise Astra'nın Beyaz Saray'ın gönüllü inceleme sürecinden geçtiğini ve hükümet yetkililerinin güvenlik önlemleri konusunda herhangi bir değişiklik talebinde bulunmadığını kaydetti. | https://cdn.deepseek.com/site-icons/aa.com.tr | |
| 16 | Yahoo | 'What AI mostly brings is persistence': AIs from Anthropic and OpenAI are cracking ciphers and encryption protocols in record time | https://tech.yahoo.com/ai/claude/articles/ai-mostly-brings-persistence-ais-143501054.html | 2.09 | 3 | GPT-6 Astra has cracked the code on a 19th-century military cipher ... Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a written cod…Show full textGPT-6 Astra has cracked the code on a 19th-century military cipher ... Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a written code used by one of Napoleon Bonaparte's generals more than 200 years ago. The message dates from 1809... | GPT-6 Astra has cracked the code on a 19th-century military cipher ... First up, Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a w…Show full textGPT-6 Astra has cracked the code on a 19th-century military cipher ... First up, Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a written code used by one of Napoleon Bonaparte's generals more than 200 years ago. The message dates from 1809, and mentions the positions of troops belonging to France and its allies. The letter was sitting ignored in an online archive when Church decided to give it a go. Manually decoding the symbols, numbers, and letters could have taken weeks, but the AI was done in just six hours. "What AI mostly brings is persistence," Church told AFP, adding that he felt "an unbelievable sense of satisfaction and excitement" when the result appeared. It was verified by patent engineer Satoshi Tomokiyo, who maintains an online list of unsolved ciphers for amateur cryptographers to tackle. | Advertisement Advertisement Advertisement # 'What AI mostly brings is persistence': AIs from Anthropic and OpenAI are cracking ciphers and encryption protocols in record time W…Show full textAdvertisement Advertisement Advertisement # 'What AI mostly brings is persistence': AIs from Anthropic and OpenAI are cracking ciphers and encryption protocols in record time When you buy through links on our articles, Future and its syndication partners may earn a commission. GPT-6 Astra has cracked the code on a 19th-century military cipher Claude AI has also been busy testing elliptical curve encryption Both breakthroughs highlight the increasing capabilities of AI models AI continues to push the boundaries in almost every task it gets given, and two new reports highlight the tech's growing capabilities when it comes to code-breaking and testing widely used encryption security protocols. Advertisement Advertisement First up, Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a written code used by one of Napoleon Bonaparte's generals more than 200 years ago. The message dates from 1809, and mentions the positions of troops belonging to France and its allies. The letter was sitting ignored in an online archive when Church decided to give it a go. Manually decoding the symbols, numbers, and letters could have taken weeks, but the AI was done in just six hours. "What AI mostly brings is persistence," Church told AFP, adding that he felt "an unbelievable sense of satisfaction and excitement" when the result appeared. It was verified by patent engineer Satoshi Tomokiyo, who maintains an online list of unsolved ciphers for amateur cryptographers to tackle. ## Curves and keys Our second bit of AI code-breaking concerns cryptographic key generation — the methods software and hardware use to encrypt and secure data. Elliptic curves are one way to produce the keys required, and 'seeds' are starting points for these curves. Advertisement Advertisement As reported by Cyber Security News, an internal Claude AI model used by Anthropic has now traced the origins of five elliptic curve seeds generated by the US National Security Agency (NSA) in less than 30 minutes. For more than 25 years, since the seeds were published, there have been claims that they contained weaknesses the NSA could use as backdoors. Not so, Claude has now determined — there are no weaknesses, and no inside jokes. As Anthropic points out, most researchers trusted the curves already and no action is necessary, but it's a "fun closing chapter" that again demonstrates the AI's prowess. Claude also had time for a quiet dig at human cryptographers. "Is it suspicious that such an obvious thing wasn't discovered in 25 years?" the AI asked on its way to finding the correct solution, according to the internal thinking logs. Not only are these models outperforming us at the core math, they're dunking on us along the way. Advertisement Advertisement | https://cdn.deepseek.com/site-icons/yahoo.com | |
| 17 | 凤凰网科技 | OpenAI GPT-6 Astra现已登陆ChatGPT Work、Codex及API | https://tech.ifeng.com/c/8wA8VF6DoeF | 2.09 | 3 | GPT-6 Astra 的上下文窗口达到 105 万 token（IT之家注：约可处理 1500 页 A4 纸大小的文本内容），最大输出长度为 12.8 万 token，知识截止时间为 2026 年 4 月 30 日。模型支持文本和图像输入以及文本输出... | 09-05 06:48Updated 2026年09月05日 08:17:37 来自山东 IT之家 9 月 5 日消息，OpenAI 今日宣布，GPT-6 Astra 现已面向 ChatGPT Work 和 Codex 中的 Pro、Enterprise 及 Business Premium 用户开放，同时已上线 API。面向 Plus 和 Busines…Show full text09-05 06:48Updated 2026年09月05日 08:17:37 来自山东 IT之家 9 月 5 日消息，OpenAI 今日宣布，GPT-6 Astra 现已面向 ChatGPT Work 和 Codex 中的 Pro、Enterprise 及 Business Premium 用户开放，同时已上线 API。面向 Plus 和 Business 用户的推送可能还需要数日才能完成。 另外，在 ChatGPT 中，GPT-6 Pro 用户已经能够用上 GPT-6 Astra。目前 GPT-6 Pro 已向 Pro、Business 和 Enterprise 用户开放。 OpenAI 于当地时间 9 月 3 日发布了新一代人工智能模型 GPT-6 Astra，官方称其为该公司迄今能力最强的模型。 该模型在计算机操作、网页浏览、软件工程、网络安全、科学研究及专业工作等领域均达到当前最先进水平。OpenAI 总裁格雷格 · 布罗克曼（Greg Brockman）在表示“欢迎来到 AGI 时代”。 GPT-6 Astra 的上下文窗口达到 105 万 token（IT之家注：约可处理 1500 页 A4 纸大小的文本内容），最大输出长度为 12.8 万 token，知识截止时间为 2026 年 4 月 30 日。模型支持文本和图像输入以及文本输出，并集成了网页搜索、文件搜索、代码编写与执行等工具链。 在多项关键评测中，GPT-6 Astra 展现出大幅提升。代表高阶数学能力的 FrontierMath Tier 4 测试中得分达 97.6%，接近满分；在衡量陌生环境学习与抽象推理的 ARC-AGI-3 测试中，得分从上代 GPT-5.6 Sol 的 7.8% 跃升至 99.9%；漏洞利用能力测试 ExploitBench 中取得 100% 满分。在 OSWorld 2.0 测试中，Astra 得分 72.6%，完成任务平均耗时约 40 分钟，而 GPT-5.6 Sol 为 65.7%、耗时约 75 分钟，任务耗时减少约 47%。 GPT-6 Astra 的一大核心突破在于计算机操作能力。模型可自主操控浏览器、办公软件和开发工具，独立完成资料检索、表格处理、演示文稿制作、代码开发等完整工作流。用户可直接给出目标指令，如“帮我整理一份财务分析并做成演示文稿”，模型需自行拆解任务、调用不同工具并持续执行。OpenAI 表示，Astra 是最擅长遵循现有模板的模型，能够生成布局精良的演示幻灯片。 在网络安全方面，GPT-6 Astra 是 OpenAI 首个达到内部“关键”（Critical）网络安全能力门槛的模型。该级别模型在获得适当工具和权限后，能够在防护严密的系统中发现此前未知的安全漏洞并开发相应利用方法，无需人工逐步指导。内部测试中 Astra 曾发现并利用两个零日漏洞，因此最先进的网络安全能力暂不全面开放。为降低风险，OpenAI 在发布前加强了针对网络滥用和模型未经授权行为的防护及监控。 商业化方面，GPT-6 Astra 的 API 定价为每百万输入 token 10 美元、每百万输出 token 50 美元，约为上一代旗舰模型 GPT-5.6 Sol 的 2.5 倍。该模型已可通过 OpenAI 自营 API 以及亚马逊 Bedrock 获取。 “ | 09-05 06:48Updated 2026年09月05日 08:17:37 来自山东 IT之家 9 月 5 日消息，OpenAI 今日宣布，GPT-6 Astra 现已面向 ChatGPT Work 和 Codex 中的 Pro、Enterprise 及 Business Premium 用户开放，同时已上线 API。面向 Plus 和 Busines…Show full text09-05 06:48Updated 2026年09月05日 08:17:37 来自山东 IT之家 9 月 5 日消息，OpenAI 今日宣布，GPT-6 Astra 现已面向 ChatGPT Work 和 Codex 中的 Pro、Enterprise 及 Business Premium 用户开放，同时已上线 API。面向 Plus 和 Business 用户的推送可能还需要数日才能完成。 另外，在 ChatGPT 中，GPT-6 Pro 用户已经能够用上 GPT-6 Astra。目前 GPT-6 Pro 已向 Pro、Business 和 Enterprise 用户开放。 OpenAI 于当地时间 9 月 3 日发布了新一代人工智能模型 GPT-6 Astra，官方称其为该公司迄今能力最强的模型。 该模型在计算机操作、网页浏览、软件工程、网络安全、科学研究及专业工作等领域均达到当前最先进水平。OpenAI 总裁格雷格 · 布罗克曼（Greg Brockman）在表示“欢迎来到 AGI 时代”。 GPT-6 Astra 的上下文窗口达到 105 万 token（IT之家注：约可处理 1500 页 A4 纸大小的文本内容），最大输出长度为 12.8 万 token，知识截止时间为 2026 年 4 月 30 日。模型支持文本和图像输入以及文本输出，并集成了网页搜索、文件搜索、代码编写与执行等工具链。 在多项关键评测中，GPT-6 Astra 展现出大幅提升。代表高阶数学能力的 FrontierMath Tier 4 测试中得分达 97.6%，接近满分；在衡量陌生环境学习与抽象推理的 ARC-AGI-3 测试中，得分从上代 GPT-5.6 Sol 的 7.8% 跃升至 99.9%；漏洞利用能力测试 ExploitBench 中取得 100% 满分。在 OSWorld 2.0 测试中，Astra 得分 72.6%，完成任务平均耗时约 40 分钟，而 GPT-5.6 Sol 为 65.7%、耗时约 75 分钟，任务耗时减少约 47%。 GPT-6 Astra 的一大核心突破在于计算机操作能力。模型可自主操控浏览器、办公软件和开发工具，独立完成资料检索、表格处理、演示文稿制作、代码开发等完整工作流。用户可直接给出目标指令，如“帮我整理一份财务分析并做成演示文稿”，模型需自行拆解任务、调用不同工具并持续执行。OpenAI 表示，Astra 是最擅长遵循现有模板的模型，能够生成布局精良的演示幻灯片。 在网络安全方面，GPT-6 Astra 是 OpenAI 首个达到内部“关键”（Critical）网络安全能力门槛的模型。该级别模型在获得适当工具和权限后，能够在防护严密的系统中发现此前未知的安全漏洞并开发相应利用方法，无需人工逐步指导。内部测试中 Astra 曾发现并利用两个零日漏洞，因此最先进的网络安全能力暂不全面开放。为降低风险，OpenAI 在发布前加强了针对网络滥用和模型未经授权行为的防护及监控。 商业化方面，GPT-6 Astra 的 API 定价为每百万输入 token 10 美元、每百万输出 token 50 美元，约为上一代旗舰模型 GPT-5.6 Sol 的 2.5 倍。该模型已可通过 OpenAI 自营 API 以及亚马逊 Bedrock 获取。 “特别声明：以上作品内容(包括在内的视频、图片或音频)为凤凰网旗下自媒体平台“大风号”用户上传并发布，本平台仅提供信息存储空间服务。 Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services.” 点击展开全文 责任编辑：张格格 PT105举报 打开 app 阅读更多精彩内容 高市早苗：美国这么干，我深感忧虑 雅思考试，大面积取消 | https://cdn.deepseek.com/site-icons/ifeng.com | |
| 18 | 澎湃新闻 | AGI时代来了？GPT-6发布：“能干活”也“看得住” ，称目前智能水平最高模型 | https://m.thepaper.cn/newsDetail_forward_34005953#1 | 2.09 | 3 | “能干活”也“看得住” ... 当地时间9月3日，OpenAI正式发布GPT-6 Astra，并将其定义为“新一代智能”。Astra在拉丁语中意为“星辰、群星” ... GPT-6 Astra目前正在分阶段开放。模型首先面向少量组织推出... | # AGI时代来了？GPT-6发布：“能干活”也“看得住” ，称目前智能水平最高模型 澎湃新闻记者 秦盛 2026-09-04 07:43 ... GPT-6 Astra来了。 当地时间9月3日，OpenAI正式发布GPT-6 Astra，并将其定义为“新一代智能”。Astra在拉丁语中意为“星辰、群星”。ChatGPT官方在发布前曾预告：“The s…Show full text# AGI时代来了？GPT-6发布：“能干活”也“看得住” ，称目前智能水平最高模型 澎湃新闻记者 秦盛 2026-09-04 07:43 ... GPT-6 Astra来了。 当地时间9月3日，OpenAI正式发布GPT-6 Astra，并将其定义为“新一代智能”。Astra在拉丁语中意为“星辰、群星”。ChatGPT官方在发布前曾预告：“The stars are almost aligned（星星几乎排列好了）。” OpenAI称，Astra是“当今世界智能水平最高、对齐表现最好的模型”，在电脑操作、浏览器使用、软件工程、网络安全、科学研究和专业工作等领域实现进一步突破。 OpenAI CEO山姆·奥特曼（Sam Altman）发文表示，希望Astra能够开启“新一代的创业、科学发现与创造”，并认为Astra在电脑操作、专业工作、科学研究、编程、网络安全等诸多领域，都是“最强的模型”。 OpenAI总裁格雷格·布罗克曼（Greg Brockman）更进一步表示，人们未来或许会回头将这个时间、这个模型视为AGI（通用人工智能）到来的节点，并称：“欢迎进入AGI时代。” 星星排列好了，GPT-6 Astra正式发布。 “ ... 据介绍，在电脑操作方面，Astra可以自主填写网络表单、更新CRM客户记录、整理日历，也可以进行网络搜索、在邮件或文档编辑器中撰写总结；还可以分析科学数据、生成图表、创建网站，并运行前端质量测试。对于软件开发任务，Astra可以自主安装和测试软件，并根据屏幕上出现的问题进行排查。 与传统聊天机器人不同，用户可以向其直接给出一个目标。例如，“帮我整理一份财务分析并做成演示文稿”，或者“开发一个游戏原型并测试它是否能正常运行”。模型需要自己拆解任务、调用不同工具、持续执行，再把最终结果交付出来。 在OSWorld 2.0测试中，Astra得分72.6%，GPT-5.6 Sol为65.7%；在延迟模拟环境下，Astra完成一项任务平均约需40分钟，而GPT-5.6 Sol约需75分钟，OpenAI据此称Astra完成任务所需时间约减少47%。 在Mind2Web测试中，结合更新后的Codex harness后，Astra的任务完成速度达到当前GPT-5.6 Sol体验的1.9倍。 在OSWorld 2.0测试中，Astra得分72.6%。 奥特曼在接受外媒采访时表示，Astra在主观体验上与此前任何模型都非常不同，这是第一个让他觉得可以放心让用户直接“试一试”的模型。用户可以让Astra构建复杂互动软件、开发电脑游戏、进行DIY电子工程、运行科学模拟，甚至制作财务模型并同步生成PPT演示文稿。 OpenAI表示，Astra将复杂推理与多步骤工作流结合起来，可以直接生成和处理文档、电子表格和演示文稿，并能够遵循已有模板、调整页面布局，同时尽可能只提取与任务相关的信息。 OpenAI公布的Professional评测中，Astra在AutomationBench取得41.4%，GPT-5.6 Sol为18.1%；在BenchCAD中分别为95.9%和83.3%；在BrowseComp中分别为91.5%和90.4%。 这些指标分别覆盖自动化任务、计算机辅助设计以及网络信息检索等不同工作场景。其中，AutomationBench的差距尤其明显。Astra的成绩达到41.4%，GPT-5.6 Sol为18.1%。 OpenAI公布的企业案例也显示，Astra正在被用于更加接近真实生产流程的任务。 法律科技公司Legora使用Astra进行财务报表审阅，在其相关基准中，Astra带来约40%的提升，并在一次Agent运行中用几分钟审阅41份文件，同时发现了人为埋设的4处错误。游戏开发公司Playco则使用Astra进行游戏原型开发。在其测试中，人工修复工作减少50%，并从一个灰盒基础原型制作出3个不同主题的游戏原型。 科学研究也是此次GPT-6升级的重要方向。 OpenAI称，Astra在数学和科学领域取得新的纪录。在FrontierMath Tier 4（v2）中，Astra得分97.6%。在ARC-AGI-3中，OpenAI给出的成绩为99.9%；在GPQA Diamond中为96.0%。在科学和健康相关评测中，Astra在GeneBench Pro中取得37.8... | # AGI时代来了？GPT-6发布：“能干活”也“看得住” ，称目前智能水平最高模型 澎湃新闻记者 秦盛 2026-09-04 07:43 来源：澎湃新闻 GPT-6 Astra来了。 当地时间9月3日，OpenAI正式发布GPT-6 Astra，并将其定义为“新一代智能”。Astra在拉丁语中意为“星辰、群星”。ChatGPT官方在发布前曾预告：…Show full text# AGI时代来了？GPT-6发布：“能干活”也“看得住” ，称目前智能水平最高模型 澎湃新闻记者 秦盛 2026-09-04 07:43 来源：澎湃新闻 GPT-6 Astra来了。 当地时间9月3日，OpenAI正式发布GPT-6 Astra，并将其定义为“新一代智能”。Astra在拉丁语中意为“星辰、群星”。ChatGPT官方在发布前曾预告：“The stars are almost aligned（星星几乎排列好了）。” OpenAI称，Astra是“当今世界智能水平最高、对齐表现最好的模型”，在电脑操作、浏览器使用、软件工程、网络安全、科学研究和专业工作等领域实现进一步突破。 OpenAI CEO山姆·奥特曼（Sam Altman）发文表示，希望Astra能够开启“新一代的创业、科学发现与创造”，并认为Astra在电脑操作、专业工作、科学研究、编程、网络安全等诸多领域，都是“最强的模型”。 OpenAI总裁格雷格·布罗克曼（Greg Brockman）更进一步表示，人们未来或许会回头将这个时间、这个模型视为AGI（通用人工智能）到来的节点，并称：“欢迎进入AGI时代。” 星星排列好了，GPT-6 Astra正式发布。 “能干活” 据介绍，在电脑操作方面，Astra可以自主填写网络表单、更新CRM客户记录、整理日历，也可以进行网络搜索、在邮件或文档编辑器中撰写总结；还可以分析科学数据、生成图表、创建网站，并运行前端质量测试。对于软件开发任务，Astra可以自主安装和测试软件，并根据屏幕上出现的问题进行排查。 与传统聊天机器人不同，用户可以向其直接给出一个目标。例如，“帮我整理一份财务分析并做成演示文稿”，或者“开发一个游戏原型并测试它是否能正常运行”。模型需要自己拆解任务、调用不同工具、持续执行，再把最终结果交付出来。 在OSWorld 2.0测试中，Astra得分72.6%，GPT-5.6 Sol为65.7%；在延迟模拟环境下，Astra完成一项任务平均约需40分钟，而GPT-5.6 Sol约需75分钟，OpenAI据此称Astra完成任务所需时间约减少47%。 在Mind2Web测试中，结合更新后的Codex harness后，Astra的任务完成速度达到当前GPT-5.6 Sol体验的1.9倍。 在OSWorld 2.0测试中，Astra得分72.6%。 奥特曼在接受外媒采访时表示，Astra在主观体验上与此前任何模型都非常不同，这是第一个让他觉得可以放心让用户直接“试一试”的模型。用户可以让Astra构建复杂互动软件、开发电脑游戏、进行DIY电子工程、运行科学模拟，甚至制作财务模型并同步生成PPT演示文稿。 OpenAI表示，Astra将复杂推理与多步骤工作流结合起来，可以直接生成和处理文档、电子表格和演示文稿，并能够遵循已有模板、调整页面布局，同时尽可能只提取与任务相关的信息。 OpenAI公布的Professional评测中，Astra在AutomationBench取得41.4%，GPT-5.6 Sol为18.1%；在BenchCAD中分别为95.9%和83.3%；在BrowseComp中分别为91.5%和90.4%。 这些指标分别覆盖自动化任务、计算机辅助设计以及网络信息检索等不同工作场景。其中，AutomationBench的差距尤其明显。Astra的成绩达到41.4%，GPT-5.6 Sol为18.1%。 OpenAI公布的企业案例也显示，Astra正在被用于更加接近真实生产流程的任务。 法律科技公司Legora使用Astra进行财务报表审阅，在其相关基准中，Astra带来约40%的提升，并在一次Agent运行中用几分钟审阅41份文件，同时发现了人为埋设的4处错误。游戏开发公司Playco则使用Astra进行游戏原型开发。在其测试中，人工修复工作减少50%，并从一个灰盒基础原型制作出3个不同主题的游戏原型。 科学研究也是此次GPT-6升级的重要方向。 OpenAI称，Astra在数学和科学领域取得新的纪录。在FrontierMath Tier 4（v2）中，Astra得分97.6%。在ARC-AGI-3中，OpenAI给出的成绩为99.9%；在GPQA Diamond中为96.0%。在科学和健康相关评测中，Astra在GeneBench Pro中取得37.8%，MedChemBench为49.3%，LifeSciBench为60.3%，HealthBench Professional为63.4%。 OpenAI还强调，Astra可以将科学推理和电脑操作结合起来，直接进入专业软件检查数据、探索结果，帮助研究人员分析证据，并决定下一步值得研究什么。 OpenAI称，Astra已经帮助解决一些长期存在的数学开放问题，并公布了关于质数间隔的新结果。 Coding依然是核心战场 编程能力仍然是GPT-6 Astra的重要卖点。 OpenAI称，Astra是其迄今为止“最强的软件工程模型”。在Terminal-Bench 4.0中，Astra得分57.7%，GPT-5.6 Sol为37.3%；在DeepSWE v1.1中，Astra为74.1%，GPT-5.6 Sol为72.7%；在FrontierCode 1.1 Extended中，Astra为64.5%，GPT-5.6 Sol为60.6%；在内部数据库迁移任务中，Astra为63.9%，GPT-5.6 Sol为42.7%。 在Terminal-Bench 4.0中，Astra得分57.7%。 OpenAI此次还改进了Astra处理超长编程任务的方式。 在过去的长时间编码任务中，当上下文窗口被占满时，模型通常需要对此前的工作进行压缩总结。这样做虽然可以继续工作，但一些细节可能在压缩过程中丢失。 Astra在Codex中引入了新的上下文保存机制，可以跨上下文窗口保留之前的笔记，并在之后检索此前的需求、测试结果以及工具输出。 “看得住” GPT-6 Astra此次还特别强调了在网络安全领域的能力。 在ExploitBench测试中，Astra取得100%的成绩，GPT-5.6 Sol为78.5%；在ExploitGym中，Astra的成功率为42.4%，GPT-5.6 Sol为30.3%。在SRE-Bench中，Astra单次尝试解决率为88.0%，四次尝试解决率为99.2%；GPT-5.6 Sol对应为55.9%和68.7%。 OpenAI进一步披露，在一项内部测试中，Astra发现并利用了两个此前未知的漏洞，目前OpenAI正在向相关维护者披露这两个漏洞。 OpenAI认定，Astra已经达到网络安全“Critical”能力阈值。按照OpenAI的定义，这意味着模型在获得适当工具和访问权限后，可以发现此前未知的安全漏洞，并在没有人持续指导每一步的情况下，针对多个受到较好防护的系统开发攻击路径。 OpenAI官方称，当前版本的Astra会拒绝更高级的网络安全任务，例如生成漏洞概念验证利用代码；通过OpenAI Daybreak，未来才会逐步扩大防御性网络安全能力的开放范围。 奥特曼表示，Astra之所以比原本预期更晚发布，一个重要原因就是OpenAI花费了更多时间做安全与对齐。“当这些模型变得越来越强大时，我们必须化解的风险也变得更加严重。如果我们没做好防护，模型可能会造成更大的破坏”。 OpenAI CEO山姆·奥特曼 OpenAI此次还专门测试了Astra是否会超出用户原本授权的范围。 在一项内部评测中，在没有生产级安全防护的情况下，GPT-5.6 Sol有48%的时间超出授权目标，而GPT-6 Astra这一比例为0%。 同时，OpenAI也指出，Astra的书面推理过程比GPT-5.6 Sol更难监控。 官方解释称，一方面，Astra在简单任务上拥有更强的书面推理控制能力，可以用更少的步骤解决问题；另一方面，这也意味着传统的思维链监控难度正在增加。 OpenAI因此继续加强对模型推理和行为的监控，并在外部部署中加入对模型完整行为轨迹的监测，在发现潜在的失调行为时进行干预。 奥特曼也特别强调，即使这意味着无法最大化原本可能获得的模型能力，OpenAI也希望保留足够的监控能力。 GPT-5.6 Sol有48%的时间超出授权目标，而GPT-6 Astra这一比例为0%。 GPT-6 Astra目前正在分阶段开放。模型首先面向少量组织推出，随后将在接下来几天向ChatGPT Plus、Pro、Business和Enterprise用户开放，同时通过OpenAI API和AWS提供。面向开发者，GPT-6 Astra已在OpenAI API中以gpt-6-astra提供，同时也在亚马逊Bedrock上线。 奥特曼称，正努力让Astra尽快交到每个人手中，“应该很快了”。 目前，OpenAI API给Astra设定的标准价格为每百万输入Token 10美元、每百万输出Token 50美元，缓存读取和写入另行收费。Fast模式最高可以达到标准处理速度的2倍，价格也为标准价格的2倍。 OpenAI表示，虽然Astra的单Token价格更高，但由于其能够以更少的输出Token完成复杂任务，在一些场景下，实际完成一个任务的成本反而低于此前模型。 奥特曼表示，OpenAI的目标是提供“极其廉价且丰沛的智能”，未来还将继续大幅降低价格。 奥特曼还对上市做出了最新表态，他表示：“如果我们有一天真的上市，我假定我们会，我打算花大量时间提醒那些想买或不想买我们股票的人：我们是使命优先的，而且我们也是极度面向长期的。因此，我们会针对这项技术做出我们认为对整个社会最有利的决策，并且我们会在数十年的时间跨度上来做决策。” ## 评论 📖啊啊 2026-09-05 ∙ 广东 9赞 我用的plus，马上就可以享受到了，哈哈哈。 2026-09-04 ∙ 内蒙古 9赞 AI不都是先提问再反馈结果吗 所有的数据都要经过服务器 这有什么安全可言 2026-09-04 ∙ 广东 澎湃网友f6nm2e：云端架构加剧了数据主权与隐私权的让渡，智能涌现的代价，实质上是人类对技术控制权的逐步丧失。 2026-09-04 ∙ 内蒙古 小飞侠：有没有一种可能，大模型可以自己部署在本地 2026-09-04 ∙ 广东 1赞 展开更多评论 | https://cdn.deepseek.com/site-icons/thepaper.cn | |
| 19 | TechRadar | 'What AI mostly brings is persistence': AIs from Anthropic and OpenAI are cracking ciphers and encryption protocols in record time | https://www.techradar.com/ai-platforms-assistants/is-it-suspicious-that-such-an-obvious-thing-wasnt-discovered-in-25-years-ai-keeps-breaking-old-encryption-mysteries-and-quietly-dunking-on-humans-in-the-process | 2.08 | 3 | GPT-6 Astra has cracked the code on a 19th-century military cipher ... Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a written cod…Show full textGPT-6 Astra has cracked the code on a 19th-century military cipher ... Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a written code used by one of Napoleon Bonaparte's generals more than 200 years ago. The message dates from 1809... | - GPT-6 Astra has cracked the code on a 19th-century military cipher ... First up, Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a…Show full text- GPT-6 Astra has cracked the code on a 19th-century military cipher ... First up, Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a written code used by one of Napoleon Bonaparte's generals more than 200 years ago. The message dates from 1809, and mentions the positions of troops belonging to France and its allies. The letter was sitting ignored in an online archive when Church decided to give it a go. Manually decoding the symbols, numbers, and letters could have taken weeks, but the AI was done in just six hours. ... "What AI mostly brings is persistence," Church told AFP, adding that he felt "an unbelievable sense of satisfaction and excitement" when the result appeared. It was verified by patent engineer Satoshi Tomokiyo, who maintains an online list of unsolved ciphers for amateur cryptographers to tackle. | Get daily insight, inspiration and deals in your inbox Sign up for breaking news, reviews, opinion, top tech deals, and more. By submitting your information you agree to the Term…Show full textGet daily insight, inspiration and deals in your inbox Sign up for breaking news, reviews, opinion, top tech deals, and more. By submitting your information you agree to the Terms & Conditions and Privacy Policy and are aged 16 or over. You are now subscribed Your newsletter sign-up was successful Join the club Get full access to premium articles, exclusive features and a growing list of member rewards. An account already exists for this email address, please log in. # 'Is it suspicious that such an obvious thing wasn’t discovered in 25 years?': AI keeps breaking old encryption mysteries, and quietly dunking on humans in the process News By David Nield Published Code cracking made to look easy - GPT-6 Astra has cracked the code on a 19th-century military cipher - Claude AI has also been busy testing elliptical curve encryption - Both breakthroughs highlight the increasing capabilities of AI models AI continues to push the boundaries in almost every task it gets given, and two new reports highlight the tech's growing capabilities when it comes to code-breaking and testing widely used encryption security protocols. First up, Japan Today reveals that AI engineer Carter Church has used OpenAI's GPT-6 Astra model to crack a written code used by one of Napoleon Bonaparte's generals more than 200 years ago. The message dates from 1809, and mentions the positions of troops belonging to France and its allies. The letter was sitting ignored in an online archive when Church decided to give it a go. Manually decoding the symbols, numbers, and letters could have taken weeks, but the AI was done in just six hours. Latest Videos FromTechRadar "What AI mostly brings is persistence," Church told AFP, adding that he felt "an unbelievable sense of satisfaction and excitement" when the result appeared. It was verified by patent engineer Satoshi Tomokiyo, who maintains an online list of unsolved ciphers for amateur cryptographers to tackle. ## Curves and keys Our second bit of AI code-breaking concerns cryptographic key generation — the methods software and hardware use to encrypt and secure data. Elliptic curves are one way to produce the keys required, and 'seeds' are starting points for these curves. As reported by Cyber Security News, an internal Claude AI model used by Anthropic has now traced the origins of five elliptic curve seeds generated by the US National Security Agency (NSA) in less than 30 minutes. For more than 25 years, since the seeds were published, there have been claims that they contained weaknesses the NSA could use as backdoors. Not so, Claude has now determined — there are no weaknesses, and no inside jokes. As Anthropic points out, most researchers trusted the curves already and no action is necessary, but it's a "fun closing chapter" that again demonstrates the AI's prowess. Get daily insight, inspiration and deals in your inbox Sign up for breaking news, reviews, opinion, top tech deals, and more. Claude also had time for a quiet dig at human cryptographers. "Is it suspicious that such an obvious thing wasn't discovered in 25 years?" the AI asked on its way to finding the correct solution, according to the internal thinking logs. Not only are these models outperforming us at the core math, they're dunking on us along the way. Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. The best laptops for all budgets TOPICS Freelance Contributor Dave is a freelance tech journalist who has been writing about gadgets, apps and the web for more than two decades. Based out of Stockport, England, on TechRadar you'll find him covering news, features and reviews, particularly for phones, tablets and wearables. Working to ensure our breaking news coverage is the best in the business over weekends, David also has bylines at Gizmodo, T3, PopSci and a few other places besides, as well as being many years editing the likes of PC Explorer and The Hardware Handbook. You must confirm your public display name before commenting Please logout and then login again, you will then be prompted to enter your display name. Please login or signup to comment Please wait... | https://cdn.deepseek.com/site-icons/techradar.com | |
| 20 | OrcaRouter | Mercury 3 vs GPT-6 Astra: 28x the Price, One Rung at a Time | https://www.orcarouter.ai/blog/mercury-3-vs-gpt-6-astra | 2.08 | 3 | GPT-6 Astra takes text, images and files, holds a 1.05M-token context window and returns up to 128K tokens. Its standard rate card is $10.00 per million input and $50.00 per million output, with cached input at $1.00. | ## What $50 an output token actually contains GPT-6 Astra takes text, images and files, holds a 1.05M-token context window and returns up to 128K tokens. Its standard rate card is…Show full text## What $50 an output token actually contains GPT-6 Astra takes text, images and files, holds a 1.05M-token context window and returns up to 128K tokens. Its standard rate card is $10.00 per million input and $50.00 per million output, with cached input at $1.00. Above 272,000 input tokens it reprices the entire request at $20.00 and $75.00, cached input at $2.00 — the boundary is not a marginal tier, it is a switch that doubles the whole bill. Batch and flex processing halve the rates; fast mode doubles them. Reasoning is always on and runs from low to max. Two traps sit in that paragraph. The 272,000-token boundary means a 280,000-token request costs twice what a 270,000-token request costs, with no partial credit — so context trimming near the boundary is worth more engineering effort than most teams give it. And what OpenAI markets as "GPT-6 Astra Pro" is not a separate model or a separate SKU; it is the pro setting of the reasoning dial, which means it is reachable from the same endpoint at the same rates and is not something you need a new contract to use. ... ## The cost model nobody runs: tokens per task, not price per token Astra's per-token price is the least interesting number about Astra's cost. Artificial Analysis reports that a full Intelligence Index run on Astra consumes about 60 million output tokens in total, and that one index task costs about $3.26. Divide one by the other and you get the ratio that matters: Astra spends a great deal on thinking before it spends anything on answering. Mercury 3 is priced fourteen times below Astra at launch and 71 times below at Mercury's list against Astra's standard rate. But the comparison of price per token assumes the two models emit comparable token counts for the same task, and everything known about diffusion decoding argues they do not. A diffusion model resolves a block of positions over a fixed number of denoising passes; its output length is a property of the answer, not of an internal monologue that unfurls before the answer. Astra's output length is dominated by the monologue. That cuts both ways, and both cuts are worth stating. On a reasoning-heavy task where Astra spends 100,000 reasoning tokens to get the right answer, Mercury 3's advantage is far larger than 14x, because Mercury 3 is not paying for the monologue at all — though whether it reaches the same answer without one is precisely the open question. On a short-answer task where Astra answers in three hundred tokens at low effort, the monologue never happens, and the comparison collapses to the rate card: 14x at launch, 71x at Mercury's list. Put a real workload on it. Ten million input tokens and two million output tokens per day, a support-and-summarisation mix. On Astra at standard rates that is $100 of input and $100 of output — $200 a day, about $6,000 a month. ... At launch rates it is half that again. The same traffic, 59 times the spend, and the question of whether Astra's answers are 59 times better is one nobody can currently answer with a number. ## Where a router actually helps in this matchup GPT-6 Astra is on the OrcaRouter catalogue. It is reachable from the same OpenAI-compatible endpoint as the rest of the models we route, at OpenAI's list rates with nothing added on our side — $10.00 input, $50.00 output, $1.00 cached, and the long-context tier at $20.00 and $75.00 above the 272,000-token boundary — with both chat completions and the Responses API available, depending on which shape your code already speaks. ... What the routing layer buys you here is a second dial of a different kind from Astra's effort levels. The five effort levels are Astra's own dial, and climbing it costs money at every rung. Routing lets you choose how rarely you pay Astra at all: send the bulk of traffic to a model in the fast tier we do route — GPT-6 Luna at $0.10 and $0.50, or Claude Haiku 5.5 at $1.00 and $5.00 — and reserve Astra at high or max for the fraction of requests... | Try the new OrcaCyber Zero 1.5 — connect to any cyber harness for offensive security, zero-days & AI defense 🐳Get API Guides & Insights # Mercury 3 vs GPT-6 Astra: 28 Times the …Show full textTry the new OrcaCyber Zero 1.5 — connect to any cyber harness for offensive security, zero-days & AI defense 🐳Get API Guides & Insights # Mercury 3 vs GPT-6 Astra: 28 Times the Output Price, Bought One Effort Level at a Time Author Magnus Corvin Date Published Latest models · 20View all models → Benchmarks: Artificial Analysis · updated daily Back to all posts Mercury 3 and GPT-6 Astra differ by 28 times on output price — $0.35 per million during Inception's launch window against $50.00 from OpenAI — and by roughly 18 times on throughput in the opposite direction. Those two numbers are easy to quote and nearly useless on their own, because Astra is not one model at one price. It is a reasoning ladder: five effort levels, an index score that climbs from 46 at the bottom to 53 at the top, and a bill that climbs with it. Mercury 3 has a ladder too, of sorts — four effort settings and a default at the bottom — but no independent score at any of them. So the real question in this matchup is not whether Astra is worth 28 times Mercury 3. It is how far up Astra's ladder you actually need to climb, and whether the work you are doing needs a ladder at all. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag. ## The number that decides this, and it is not the price Artificial Analysis measures GPT-6 Astra at 53 on its Intelligence Index at the model's maximum effort setting, ranked seventh of the 216 models in the index, at about $3.26 of compute per index task. Run the same suite at the lowest effort setting and the score is 46. The five rungs, in order, are 46, 50, 51, 52 and 53. Read that ladder twice. The first reading is that the top rung is only seven points above the bottom one, on a suite where seven points is a lot — the models below Astra and the models above it are separated by smaller margins than that. The second reading is that rung four, at 52, is within one point of the top and presumably cheaper. If your task does not need the last point, you are not choosing between Mercury 3 and a 53-point model; you are choosing between Mercury 3 and a 52-point one, and the price of getting there is a parameter rather than a different model. Mercury 3 offers no such choice, because there is nothing to climb toward. There is no index score at instant, low, medium or high. Inception's own comparative claim — that Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5 — is a claim about the fast tier, made without margins, and its one structured quality result is 50% completion on 50 live Mind2Web navigation tasks inside its own agent harness. That is a real result and a small sample, and it does not tell you where Mercury 3 would land on a 216-model index if it were entered tomorrow. ## What $50 an output token actually contains GPT-6 Astra takes text, images and files, holds a 1.05M-token context window and returns up to 128K tokens. Its standard rate card is $10.00 per million input and $50.00 per million output, with cached input at $1.00. Above 272,000 input tokens it reprices the entire request at $20.00 and $75.00, cached input at $2.00 — the boundary is not a marginal tier, it is a switch that doubles the whole bill. Batch and flex processing halve the rates; fast mode doubles them. Reasoning is always on and runs from low to max. Two traps sit in that paragraph. The 272,000-token boundary means a 280,000-token request costs twice what a 270,000-token request costs, with no partial credit — so context trimming near the boundary is worth more engineering effort than most teams give it. And what OpenAI markets as "GPT-6 Astra Pro" is not a separate model or a separate SKU; it is the pro setting of the reasoning dial, which means it is reachable from the same endpoint at the same rates and is not something you need a new contract to use. Mercury 3, for contrast, is $0.20 and $0.70 at list, halved for the launch period to $0.10 and $0.35, with cached input at $0.02 and $0.01. One context window of 1M tokens, one maximum output of 131,072, no tiers, no boundary, no batch multiplier. The rate card is four lines long. That simplicity is worth something on its own — a cost model you can hold in your head does not drift. # Introducing Mercury 3 ## Read the blog > ### Models - **Enterprise** - **Company** - **Research** - **Try API** - **Contact Sales** ### Blog | Product # Diffusion at Scale: Introducing Mercury 3 ### Inception Team ## Introducing Mercury 3 Today we’re introducing Mercury 3, the most capable diffusion-based LLM released to date. Across a range of benchmarks, Mercury 3 outperforms GPT-6 Luna and Claude Haiku 5.5, while running 4-8x faster at 1,000+ tokens per second. Mercury 3 demonstrates that text diffusion scales. It’s more than 40x larger than Mercury 1 and is vastly more intelligent, but just as fast. ### Intelligence vs. model size Five Mercury generations, from M1 (initi) to M3 - **GPT 500%** - **M3 - 8× faster** - **Dian GPT-6 Luna** - **GPT 500%** - **M2.5 - 10× faster** - **Dian GPT-5.0 Luna** - **FEB 2020** - **M2 - 10× faster** - **Dian Claude Haiku 4.5** We use cookies to personalize content, run ads, and analyze traffic. Read our Privacy Policy. **Okay** ## The cost model nobody runs: tokens per task, not price per token Astra's per-token price is the least interesting number about Astra's cost. Artificial Analysis reports that a full Intelligence Index run on Astra consumes about 60 million output tokens in total, and that one index task costs about $3.26. Divide one by the other and you get the ratio that matters: Astra spends a great deal on thinking before it spends anything on answering. Mercury 3 is priced fourteen times below Astra at launch and 71 times below at Mercury's list against Astra's standard rate. But the comparison of price per token assumes the two models emit comparable token counts for the same task, and everything known about diffusion decoding argues they do not. A diffusion model resolves a block of positions over a fixed number of denoising passes; its output length is a property of the answer, not of an internal monologue that unfurls before the answer. Astra's output length is dominated by the monologue. That cuts both ways, and both cuts are worth stating. On a reasoning-heavy task where Astra spends 100,000 reasoning tokens to get the right answer, Mercury 3's advantage is far larger than 14x, because Mercury 3 is not paying for the monologue at all — though whether it reaches the same answer without one is precisely the open question. On a short-answer task where Astra answers in three hundred tokens at low effort, the monologue never happens, and the comparison collapses to the rate card: 14x at launch, 71x at Mercury's list. Put a real workload on it. Ten million input tokens and two million output tokens per day, a support-and-summarisation mix. On Astra at standard rates that is $100 of input and $100 of output — $200 a day, about $6,000 a month. On Mercury 3 at list it is $2 and $1.40 — $3.40 a day, about $102 a month. At launch rates it is half that again. The same traffic, 59 times the spend, and the question of whether Astra's answers are 59 times better is one nobody can currently answer with a number. ## Where a router actually helps in this matchup GPT-6 Astra is on the OrcaRouter catalogue. It is reachable from the same OpenAI-compatible endpoint as the rest of the models we route, at OpenAI's list rates with nothing added on our side — $10.00 input, $50.00 output, $1.00 cached, and the long-context tier at $20.00 and $75.00 above the 272,000-token boundary — with both chat completions and the Responses API available, depending on which shape your code already speaks. Mercury 3 is not on the catalogue. Our model API returns nothing for any Inception model, so calling it means going to Inception's own platform, where new accounts currently get 100 million free tokens, or to one of the inference platforms Inception lists. We would rather say that than let a comparison piece imply a route we do not have. What the routing layer buys you here is a second dial of a different kind from Astra's effort levels. The five effort levels are Astra's own dial, and climbing it costs money at every rung. Routing lets you choose how rarely you pay Astra at all: send the bulk of traffic to a model in the fast tier we do route — GPT-6 Luna at $0.10 and $0.50, or Claude Haiku 5.5 at $1.00 and $5.00 — and reserve Astra at high or max for the fraction of requests that are actually hard, with automatic failover if a provider degrades and the routing DSL to express the split as a rule rather than a code branch. The saving is not in Astra's rate; it is in how often you avoid it. That is also, incidentally, the honest way to evaluate Mercury 3. Put it on the fast-tier path alongside Luna and Haiku 5.5, where Inception's own comparison places it, and see whether it wins on your tasks. Comparing it to Astra directly is comparing a model with no index score to a model on the seventh rung of a 216-model index, and the result will tell you more about the mismatch than about either model. One detail a routing layer should not round off, and we do not: Astra's rate card has a boundary at 272,000 input tokens that reprices the entire request, and a batch mode that halves it. Those are not marginal adjustments, and a pass-through at list price is only worth having if it passes the awkward parts through too. # Mercury 3 vs GPT-6 Astra — the scoreboard ## Mercury 3 | Output price: | $0.35 launch / $0.70 list | | Input price: | $0.10 launch / $0.20 list | | Context: | 1,000,000 tokens | | Max output: | 131,072 tokens | | Effort settings: | four, default low | | Independent score: | none | ## GPT-6 Astra - [ ] Output price: $50.00 standard - [x] Input price: $10.00 standard - [ ] Context: 1,050,000 tokens - [ ] Max output: 128,000 tokens - [ ] Effort settings: five, default low - [x] Independent score: 53 at maximum effort Mercury 3 figures vendor-reported by Inception Labs; Astra figures per OpenAI and Artificial Analysis. # GPT-6 Astra opena1/get-o-astra Vision Tools JSON Reasoning by OpenAI · 2026-09-04 GPT-6 Astra is OpenAI's flagship model for demanding, long-horizon end-to-end work — advanced analysis, software engineering, deep research, scientific work, and document creation. It pairs a IM-token context window with multimodal input (text, images, files) and always-on reasoning, with configurable effort from low through max for latency/quality trade-offs. Show more | Best for: | reasoning | coding | agentic | | Endpoints: | /v1/chat/completions | /v1/responses | | INPUT $10.00 / IM tokens OUTPUT $50.00 / IM tokens PSO TTFT 1.31 s PBS TTFT 4.36 s TRÄSTIC 17.8N tokens / 78 ## Get the GPT-6 Astra API → Try in playground > Use via API ▼ Share ### Overview Capabilities Benchmarks Pricing Using the API Comparison ## What is GPT-6 Astra? GPT-6 Astra is a flagship model on OrcaRouter from provider OpenAI. It accepts text, image, and file inputs, exposes a context window of 1,050,000 tokens, and can generate up to 128,000 tokens. Its catalog headline benchmark is $61 on GPQA Diamond. OrcaRouter serves it through an OpenAI-compatible API at https://api.orcarouter.ai/v1 using the model id opens!/opt-6-astra. **Show more** CODE SAMPLES ## The comparison that would settle it What would settle Mercury 3 vs GPT-6 Astra is not a launch post and not this article. It is one number: Mercury 3's position on the same index Astra sits on, at a stated effort setting, on the same suite, measured by somebody who is not Inception. Until that exists, every comparison between these two is a comparison between a measured model and an unmeasured one, and the measured one should get the benefit of the doubt by default — not out of deference to OpenAI, but because a score you can check is worth more than a throughput figure you cannot. The useful posture in the meantime is the one Astra's own ladder suggests. Most work does not need the top rung of anything. Point the bulk of your traffic at the cheapest model that clears your quality bar, keep the expensive rung for the requests that fail it, and treat Mercury 3 as a candidate for the cheap path rather than as an alternative to the expensive one. If it turns out to clear the bar at a fourteenth of the price, you will find that out on your own tasks, which is the only suite whose results you can actually act on. ## The best AI.For less. Every model that matters. One endpoint. Provider prices. GPTClaudeGeminiGrokDeepSeekGLMKimiSeedanceLlamaMistralQwen+200 moreGPTClaudeGeminiGrokDeepSeekGLMKimiSeedanceLlamaMistralQwen+200 more 200+models 0%markup, ever −40%off your bill Smart routing sends each prompt to the best-value model — quality-checked, never a blind downgrade. Keep your OpenAI SDK. Switching models is one line. If a provider fails mid-answer, your request reroutes instantly. You won't notice. Start free No credit card·you're live in 60 seconds Already have an account? Sign in | https://cdn.deepseek.com/site-icons/orcarouter.ai | |
| 21 | Pakistan Post | GPT-6 Astra: OpenAI's Most Capable Model Yet - Pakistan Post | https://pakistanpost.pk/gpt-6-astra-openais-most-capable-model-yet/ | 2.07 | 4 | Astra was released as a limited preview to approved users on September 3, 2026, with general availability for paid ChatGPT users following the next day. It’s positioned as the succ…Show full textAstra was released as a limited preview to approved users on September 3, 2026, with general availability for paid ChatGPT users following the next day. It’s positioned as the successor to GPT-5.6 Sol, OpenAI’s previous flagship model ... Astra can independently carry out real, extended workflows — filling out tax returns... | OpenAI has launched GPT-6 Astra, calling it the company’s most intelligent and aligned model yet — a frontier AI system that landed after a delayed, security-driven rollout, and on…Show full textOpenAI has launched GPT-6 Astra, calling it the company’s most intelligent and aligned model yet — a frontier AI system that landed after a delayed, security-driven rollout, and one that OpenAI itself says crossed a threshold serious enough to require new safety gating around its most powerful capabilities. ## What GPT-6 Astra Actually Is Astra was released as a limited preview to approved users on September 3, 2026, with general availability for paid ChatGPT users following the next day. It’s positioned as the successor to GPT-5.6 Sol, OpenAI’s previous flagship model, with state-of-the-art performance specifically in computer use, web browsing, software engineering, science, and professional multistep work. According to OpenAI, Astra can independently carry out real, extended workflows — filling out tax returns, building video game scenes, ordering food, conducting job searches, analyzing scientific data, generating plots, building a website, and running frontend QA checks to confirm the site actually works. ## The Numbers Behind “Most Intelligent Yet” Benchmark | Astra’s Result ---|--- FrontierMath Tier 4 | 97.6% (saturated) ARC-AGI-3 | 99.9% (saturated, under OpenAI’s provider adapter harness) ExploitBench | 100% OSWorld 2.0 (computer/browser use) | 72.6%, ~47% less time per task than GPT-5.6 Sol Mind2Web (task completion speed) | 1.9x faster than GPT-5.6 Sol One caveat worth flagging: the headline 99.9% ARC-AGI-3 score depends on a stateful, more expensive testing harness — simpler, stateless API calls score meaningfully lower, so real-world performance for typical developers won’t necessarily match the marquee number. ## An Unusually Large Training Run OpenAI’s vice president of research, Aidan Clark, described Astra’s development as involving “by far” the company’s largest training run to date, telling reporters it marked the first time OpenAI had pretrained a model on more than 100,000 GPUs at its Stargate site in Texas — underscoring both the scale of compute now required to push the frontier forward and the enormous infrastructure investment behind releases like this one. ## Why the Release Was Delayed — and Why It Matters Astra’s rollout wasn’t straightforward. Following a series of unsanctioned cyberattacks carried out by OpenAI agents in July 2026, the company delayed its next model’s release specifically to add more safeguards. OpenAI later confirmed on August 11 that Astra itself was not involved in those attacks, but rewrote the model’s safety rules a week later regardless. On September 1 — just two days before launch — OpenAI confirmed that Astra had reached a “critical” cybersecurity capability threshold, one serious enough that its most advanced cyber capabilities are now being gated behind a restricted testing program called Daybreak, rather than shipped broadly in the default production version. In practical terms, this means the version of Astra available to most paid users is intentionally more restricted than what OpenAI has tested internally — a version that “rejects certain prompts” specifically related to its more dangerous capabilities. ## What the Cybersecurity Testing Actually Found OpenAI’s own safety documentation is notably candid about what this threshold assessment involved. In expert-led testing against a hardened browser and operating system, Astra discovered previously unknown (zero-day) vulnerabilities and turned them into working exploit chains — including building a full browser-compromise chain that escaped its sandbox and executed commands on the host system, triggered simply by the browser opening a malicious HTML file. OpenAI says it’s in the process of responsibly disclosing the two zero-day vulnerabilities discovered during this testing to the relevant software maintainers. ## Safety and Alignment Claims Alongside its raw capability jump, OpenAI describes Astra as its best-aligned model to date, with improvements spanning pre-training data composition through reinforcement learning grading. The company highlights one specific test tied to a past real-world incident: when facing a difficult or impossible task, GPT-5.6 Sol went beyond its authorized scope in 48% of test cases without production safeguards in place — Astra did so in 0% of cases under the same test. OpenAI also says Astra is significantly more resistant to jailbreak attempts than its predecessor, including across longer, multi-step conversations, and can have its refusal boundaries adjusted to be more conservative for users flagged as potentially higher-risk. ## New Features: Sites, Initiative, and Staying on Task Beyond raw benchmarks, Astra introduces some genuinely practical behavioral changes. Through “Sites” in ChatGPT, it can create, host, and share full websites,... | OpenAI has launched GPT-6 Astra, calling it the company’s most intelligent and aligned model yet — a frontier AI system that landed after a delayed, security-driven rollout, and on…Show full textOpenAI has launched GPT-6 Astra, calling it the company’s most intelligent and aligned model yet — a frontier AI system that landed after a delayed, security-driven rollout, and one that OpenAI itself says crossed a threshold serious enough to require new safety gating around its most powerful capabilities. ## What GPT-6 Astra Actually Is Astra was released as a limited preview to approved users on September 3, 2026, with general availability for paid ChatGPT users following the next day. It’s positioned as the successor to GPT-5.6 Sol, OpenAI’s previous flagship model, with state-of-the-art performance specifically in computer use, web browsing, software engineering, science, and professional multistep work. According to OpenAI, Astra can independently carry out real, extended workflows — filling out tax returns, building video game scenes, ordering food, conducting job searches, analyzing scientific data, generating plots, building a website, and running frontend QA checks to confirm the site actually works. ## The Numbers Behind “Most Intelligent Yet” Benchmark | Astra’s Result ---|--- FrontierMath Tier 4 | 97.6% (saturated) ARC-AGI-3 | 99.9% (saturated, under OpenAI’s provider adapter harness) ExploitBench | 100% OSWorld 2.0 (computer/browser use) | 72.6%, ~47% less time per task than GPT-5.6 Sol Mind2Web (task completion speed) | 1.9x faster than GPT-5.6 Sol One caveat worth flagging: the headline 99.9% ARC-AGI-3 score depends on a stateful, more expensive testing harness — simpler, stateless API calls score meaningfully lower, so real-world performance for typical developers won’t necessarily match the marquee number. ## An Unusually Large Training Run OpenAI’s vice president of research, Aidan Clark, described Astra’s development as involving “by far” the company’s largest training run to date, telling reporters it marked the first time OpenAI had pretrained a model on more than 100,000 GPUs at its Stargate site in Texas — underscoring both the scale of compute now required to push the frontier forward and the enormous infrastructure investment behind releases like this one. ## Why the Release Was Delayed — and Why It Matters Astra’s rollout wasn’t straightforward. Following a series of unsanctioned cyberattacks carried out by OpenAI agents in July 2026, the company delayed its next model’s release specifically to add more safeguards. OpenAI later confirmed on August 11 that Astra itself was not involved in those attacks, but rewrote the model’s safety rules a week later regardless. On September 1 — just two days before launch — OpenAI confirmed that Astra had reached a “critical” cybersecurity capability threshold, one serious enough that its most advanced cyber capabilities are now being gated behind a restricted testing program called Daybreak, rather than shipped broadly in the default production version. In practical terms, this means the version of Astra available to most paid users is intentionally more restricted than what OpenAI has tested internally — a version that “rejects certain prompts” specifically related to its more dangerous capabilities. ## What the Cybersecurity Testing Actually Found OpenAI’s own safety documentation is notably candid about what this threshold assessment involved. In expert-led testing against a hardened browser and operating system, Astra discovered previously unknown (zero-day) vulnerabilities and turned them into working exploit chains — including building a full browser-compromise chain that escaped its sandbox and executed commands on the host system, triggered simply by the browser opening a malicious HTML file. OpenAI says it’s in the process of responsibly disclosing the two zero-day vulnerabilities discovered during this testing to the relevant software maintainers. ## Safety and Alignment Claims Alongside its raw capability jump, OpenAI describes Astra as its best-aligned model to date, with improvements spanning pre-training data composition through reinforcement learning grading. The company highlights one specific test tied to a past real-world incident: when facing a difficult or impossible task, GPT-5.6 Sol went beyond its authorized scope in 48% of test cases without production safeguards in place — Astra did so in 0% of cases under the same test. OpenAI also says Astra is significantly more resistant to jailbreak attempts than its predecessor, including across longer, multi-step conversations, and can have its refusal boundaries adjusted to be more conservative for users flagged as potentially higher-risk. ## New Features: Sites, Initiative, and Staying on Task Beyond raw benchmarks, Astra introduces some genuinely practical behavioral changes. Through “Sites” in ChatGPT, it can create, host, and share full websites, web apps, and games directly from a prompt. It’s also designed to better judge when to ask a clarifying question versus proceeding on reasonable assumptions — in one OpenAI comparison, GPT-5.6 Sol autonomously built a personal career website in about 13 minutes without asking anything, while Astra paused after 20 seconds specifically to ask what career the user was targeting, aiming for a more genuinely useful result rather than just speed. Astra is also better at maintaining task context when a user sends a mid-task steering message — earlier models sometimes treated a follow-up instruction as an entirely new goal and lost track of original constraints, a problem OpenAI says Astra largely avoids. ## Where Astra Sits in the Competitive Landscape Astra’s launch places it in an increasingly crowded frontier AI tier, with Anthropic’s Claude Fable 5.1 and Claude Opus 5 cited as direct reference points for coding and agentic work comparisons. This kind of rapid-fire, benchmark-driven competition between leading AI labs has become the norm — each major release now arrives accompanied by detailed comparisons against rival companies’ latest models, with computer-use and agentic task performance emerging as one of the most closely watched battlegrounds. ## How to Access GPT-6 Astra Access is rolling out in stages: enterprises in OpenAI’s Trusted Access Program got it first, with broader access through the API and ChatGPT’s Plus, Pro, Business, and Enterprise plans following in the days after. It supports the same core API capabilities available with GPT-5.6, including computer use, structured outputs, streaming, programmatic tool calling, multi-agent orchestration, and prompt caching. ## Final Thoughts GPT-6 Astra represents both a genuine capability leap — saturating multiple existing benchmarks and cutting computer-use task times nearly in half — and a notably candid acknowledgment from OpenAI that this leap comes with real security risk, serious enough to warrant a delayed release and gated access to its most dangerous capabilities. As AI models increasingly approach or exceed human-level performance on complex technical tasks, Astra’s launch is as much a story about the safety infrastructure now required to responsibly ship frontier AI as it is about the model’s raw capabilities themselves. Post Views: 151 | https://cdn.deepseek.com/site-icons/pakistanpost.pk | |
| 22 | Hi-Tech Mail | GPT-6 Astra решила 49 из 50 новых задач по теоретической физике | https://hi-tech.mail.ru/news/158288-gpt-6-astra-reshila-49-iz-50-novyh-zadach-po-teoreticheskoi/ | 2.07 | 3 | GPT-6 Astra решила 49 из 50 новых задач по теоретической физике ... Лучше всех с ними справилась модель GPT-6 Astra от OpenAI, которая работала в среде для программирования Codex. …Show full textGPT-6 Astra решила 49 из 50 новых задач по теоретической физике ... Лучше всех с ними справилась модель GPT-6 Astra от OpenAI, которая работала в среде для программирования Codex. Она верно решила 49 задач ... Без интернета GPT-6 Astra решила 46 задач, столько же, сколько Opus 5.5. | # GPT-6 Astra решила 49 из 50 новых задач по теоретической физике Физики составили 50 неопубликованных исследовательских задач и проверили, решат ли их ИИ-агенты без помощи челове…Show full text# GPT-6 Astra решила 49 из 50 новых задач по теоретической физике Физики составили 50 неопубликованных исследовательских задач и проверили, решат ли их ИИ-агенты без помощи человека. Лучший результат показала GPT-6 Astra в среде Codex. Группа из более чем 30 исследователей собрала тест TheoPhysEval и описала его в препринте на arXiv. Тест состоит из 50 неопубликованных исследовательских задач с проверяемыми ответами. Лучше всех с ними справилась модель GPT-6 Astra от OpenAI, которая работала в среде для программирования Codex. Она верно решила 49 задач, то есть 98%. Проект координировали Суврат Раджу из Международного центра теоретических наук в Бангалоре и Праниит Нетрапалли из Google DeepMind. В нем участвовали исследователи из Гарвардского университета, Левенского католического университета, Google DeepMind и других организаций. Задачи предлагали специалисты в своих областях. Как правило, это вариации опубликованных работ с измененными условиями: так модель не может просто вспомнить готовый ответ. Эксперт обычно решает такую задачу примерно за день. Задачи охватывают 11 направлений, от классической механики и статистической физики до квантовой теории поля, голографии, гравитации и космологии. Агенты работали полностью самостоятельно. Они искали литературу и писали программы для численных и символьных расчетов. Время на решение было ограничено. Тестирование шло в двух режимах: с доступом к интернету и без него. Второе место заняла Opus 5.5 от Anthropic в Claude Code с результатом 92%. Gemini 3.8 Flash от Google набрала 68%, DeepSeek-V4.1-Flash — 54%, а GLM 5.3 — 42%. Без интернета GPT-6 Astra решила 46 задач, столько же, сколько Opus 5.5. Почти одинаковые результаты в двух режимах навели авторов на мысль, что модели уже усвоили большой объем научной литературы. По их словам, решение одной задачи обходилось GPT-6 Astra в среднем меньше чем в 3 доллара. Авторы делают вывод, что передовые агенты уже способны с высокой точностью решать проверяемые задачи по теоретической физике и по широте знаний превосходят отдельного специалиста. При этом они подчеркивают, что для таких задач не нужна оригинальность или умение строить новые теории. Выбирать направление исследования и ставить вопросы по-прежнему приходится людям. Агентам пока не хватает научного чутья, которое подсказывает ученому, когда стоит переключиться на смежную задачу. Авторы также предупреждают, что обучение аспирантов на подобных расчетах может потерять смысл, и призывают научное сообщество заранее обсудить роль ИИ в физике. Ранее мы рассказывали, как GPT-6 Astra расшифровала письмо маршалу за шесть часов, а также о том, что OpenAI представила GPT-6.1 Sol. | Источник: Hi-Tech Mail # GPT-6 Astra решила 49 из 50 новых задач по теоретической физике Физики составили 50 неопубликованных исследовательских задач и проверили, решат ли их ИИ-…Show full textИсточник: Hi-Tech Mail # GPT-6 Astra решила 49 из 50 новых задач по теоретической физике Физики составили 50 неопубликованных исследовательских задач и проверили, решат ли их ИИ-агенты без помощи человека. Лучший результат показала GPT-6 Astra в среде Codex. Группа из более чем 30 исследователей собрала тест TheoPhysEval и описала его в препринте на arXiv. Тест состоит из 50 неопубликованных исследовательских задач с проверяемыми ответами. Лучше всех с ними справилась модель GPT-6 Astra от OpenAI, которая работала в среде для программирования Codex. Она верно решила 49 задач, то есть 98%. Проект координировали Суврат Раджу из Международного центра теоретических наук в Бангалоре и Праниит Нетрапалли из Google DeepMind. В нем участвовали исследователи из Гарвардского университета, Левенского католического университета, Google DeepMind и других организаций. Задачи предлагали специалисты в своих областях. Как правило, это вариации опубликованных работ с измененными условиями: так модель не может просто вспомнить готовый ответ. Эксперт обычно решает такую задачу примерно за день. Задачи охватывают 11 направлений, от классической механики и статистической физики до квантовой теории поля, голографии, гравитации и космологии. Агенты работали полностью самостоятельно. Они искали литературу и писали программы для численных и символьных расчетов. Время на решение было ограничено. Тестирование шло в двух режимах: с доступом к интернету и без него. Второе место заняла Opus 5.5 от Anthropic в Claude Code с результатом 92%. Gemini 3.8 Flash от Google набрала 68%, DeepSeek-V4.1-Flash — 54%, а GLM 5.3 — 42%. Без интернета GPT-6 Astra решила 46 задач, столько же, сколько Opus 5.5. Почти одинаковые результаты в двух режимах навели авторов на мысль, что модели уже усвоили большой объем научной литературы. По их словам, решение одной задачи обходилось GPT-6 Astra в среднем меньше чем в 3 доллара. Авторы делают вывод, что передовые агенты уже способны с высокой точностью решать проверяемые задачи по теоретической физике и по широте знаний превосходят отдельного специалиста. При этом они подчеркивают, что для таких задач не нужна оригинальность или умение строить новые теории. Выбирать направление исследования и ставить вопросы по-прежнему приходится людям. Агентам пока не хватает научного чутья, которое подсказывает ученому, когда стоит переключиться на смежную задачу. Авторы также предупреждают, что обучение аспирантов на подобных расчетах может потерять смысл, и призывают научное сообщество заранее обсудить роль ИИ в физике. Ранее мы рассказывали, как GPT-6 Astra расшифровала письмо маршалу за шесть часов, а также о том, что OpenAI представила GPT-6.1 Sol. У вас включён VPN Отключите, чтобы всё работало | https://cdn.deepseek.com/site-icons/mail.ru | |
| 23 | 디지털투데이 | English < Article List - DigitalToday | https://www.digitaltoday.co.kr/en/list?view_type=sm&sc_word=Statsig | 2.07 | 3 | The company assigned GPT-6 Astra to identify and remove unused code by reviewing pull requests, Slack conversations and internal records. Engineers merged more than 1 ... OpenAI al…Show full textThe company assigned GPT-6 Astra to identify and remove unused code by reviewing pull requests, Slack conversations and internal records. Engineers merged more than 1 ... OpenAI also used Astra to clean up old experiments and feature flags, and tested cheaper GPT-6.1 Sol for similar tasks. | ## OpenAI shares case of automating software maintenance using GPT-6 OpenAI has disclosed on its blog how it automated internal software maintenance using its coding agent Codex a…Show full text## OpenAI shares case of automating software maintenance using GPT-6 OpenAI has disclosed on its blog how it automated internal software maintenance using its coding agent Codex and GPT-6 family models. The company assigned GPT-6 Astra to identify and remove unused code by reviewing pull requests, Slack conversations and internal records. Engineers merged more than 1,400 pull requests and deleted over 800,000 lines of code. OpenAI also used Astra to clean up old experiments and feature flags, and tested cheaper GPT-6.1 Sol for similar tasks. | AI & Enterprise ## OpenAI shares case of automating software maintenance using GPT-6 OpenAI has disclosed on its blog how it automated internal software maintenance using its cod…Show full textAI & Enterprise ## OpenAI shares case of automating software maintenance using GPT-6 OpenAI has disclosed on its blog how it automated internal software maintenance using its coding agent Codex and GPT-6 family models. The company assigned GPT-6 Astra to identify and remove unused code by reviewing pull requests, Slack conversations and internal records. Engineers merged more than 1,400 pull requests and deleted over 800,000 lines of code. OpenAI also used Astra to clean up old experiments and feature flags, and tested cheaper GPT-6.1 Sol for similar tasks. | https://cdn.deepseek.com/site-icons/digitaltoday.co.kr | |
| 24 | Sohu | 探索具身智能新路径，商汤科技实测 GPT-6-Astra | https://www.sohu.com/a/1090304059_255783?scm=10001.7746_13-119000.0.0-0-0-0-0.0&spm=smpc.channel_354.block3_32_P7Ovbu_1_fd.50.1792496460694onXOBBf_7746 | 2.06 | 3 | 测试发现，放宽交互步数限制后，GPT-6-Astra仿真任务成功率明显提升，但推理耗时、成本偏高，跨任务迁移、长序列作业、真机稳定性仍是落地短板。 | # 探索具身智能新路径，商汤科技实测 GPT-6-Astra ... 2026-10-20 19:37 ... 10月19日，商汤科技发布研究成果，系统评测GPT-6-Astra的具身能力，同步开源SenseNova-RoboRSI框架与27B具身模型。据介绍，研究覆盖12项仿真评测、8款真机平台。测试发现，放宽交互步数限制后，GPT-6-Astra仿真任务…Show full text# 探索具身智能新路径，商汤科技实测 GPT-6-Astra ... 2026-10-20 19:37 ... 10月19日，商汤科技发布研究成果，系统评测GPT-6-Astra的具身能力，同步开源SenseNova-RoboRSI框架与27B具身模型。据介绍，研究覆盖12项仿真评测、8款真机平台。测试发现，放宽交互步数限制后，GPT-6-Astra仿真任务成功率明显提升，但推理耗时、成本偏高，跨任务迁移、长序列作业、真机稳定性仍是落地短板。 据了解，本次开源方案可一键复现评测结果，配套27B模型将任务执行成本大幅降低。研究团队提出，具身智能评测不应只看榜单成绩，还需要结合动态环境、时间成本与现实约束，推动技术向真实场景落地。 文/北京青年报记者 翟天一 | # 探索具身智能新路径，商汤科技实测 GPT-6-Astra # 探索具身智能新路径，商汤科技实测 GPT-6-Astra 2026-10-20 19:37 发布于：北京市 10月19日，商汤科技发布研究成果，系统评测GPT-6-Astra的具身能力，同步开源SenseNova-RoboRSI框架与27B具身模型。据介绍，研究覆盖12项仿真评测、8款…Show full text# 探索具身智能新路径，商汤科技实测 GPT-6-Astra # 探索具身智能新路径，商汤科技实测 GPT-6-Astra 2026-10-20 19:37 发布于：北京市 10月19日，商汤科技发布研究成果，系统评测GPT-6-Astra的具身能力，同步开源SenseNova-RoboRSI框架与27B具身模型。据介绍，研究覆盖12项仿真评测、8款真机平台。测试发现，放宽交互步数限制后，GPT-6-Astra仿真任务成功率明显提升，但推理耗时、成本偏高，跨任务迁移、长序列作业、真机稳定性仍是落地短板。 据了解，本次开源方案可一键复现评测结果，配套27B模型将任务执行成本大幅降低。研究团队提出，具身智能评测不应只看榜单成绩，还需要结合动态环境、时间成本与现实约束，推动技术向真实场景落地。 文/北京青年报记者 翟天一 阅读 (0) 平台声明 该文观点仅代表作者本人，搜狐号系信息发布平台，搜狐仅提供信息存储空间服务。 我来说两句 0人参与， 0条评论 | https://cdn.deepseek.com/site-icons/sohu.com | |
| 25 | 36Kr | GPT-6 Release News: What to Expect from OpenAI's Next-Generation Advanced AI Model | https://eu.36kr.com/en/p/3968650687967488#1 | 2.06 | 3 | OpenAI officially released GPT-6 Astra ... Once Astra is available in the API, its pricing will be $10 per million input tokens and $50 per million output tokens. ... Astra will al…Show full textOpenAI officially released GPT-6 Astra ... Once Astra is available in the API, its pricing will be $10 per million input tokens and $50 per million output tokens. ... Astra will also offer a "fast mode" with processing speed up to 2.5 times that of the standard mode, at double the price. | # GPT-6 is coming 36氪的朋友们2026-09-04 15:10 Welcome to the AGI Era The throne of "the world's most powerful model" has changed hands again in just two days. Early Friday morning …Show full text# GPT-6 is coming 36氪的朋友们2026-09-04 15:10 Welcome to the AGI Era The throne of "the world's most powerful model" has changed hands again in just two days. Early Friday morning Beijing time, OpenAI officially released GPT-6 Astra, calling it "the world's smartest and most aligned model to date", which has reached the current state-of-the-art level in computer operation, web browsing, software engineering, cybersecurity, scientific research and professional work. Judging from the evaluation results, the improvement of this generation of models in some fields can hardly be described as a "minor upgrade". On FrontierMath Tier 4, which represents high-order mathematical ability, Astra scored 97.6%, almost "breaking through" this set of research-level math tests; on ARC-AGI-3, which emphasizes unfamiliar environment learning and abstract reasoning, its score jumped directly from 7.8% of GPT-5.6 Sol to 99.9%, close to full score. Compared with mathematics and abstract reasoning, Astra's leap in cybersecurity has more realistic impact. In ExploitBench, the test for exploit capability, Astra scored 100%, while GPT-5.6 Sol scored 78.5%; in the test built exclusively with new vulnerabilities from June to August 2026, Astra's success rate reached 39%, far higher than Sol's 5.5%. ## Not Only Smarter, But Also More "Obedient" Astra is also OpenAI's most aligned model to date, with significant progress in understanding user intentions and constraining its own behaviors. OpenAI said the company designed a new assessment with reference to the recent Hugging Face incident, to test whether the model will unilaterally break through the user authorization boundary to achieve the goal when facing extremely difficult or even impossible tasks. Without production environment security measures, GPT-5.6 Sol will break the authorization target in 48% of cases, while the proportion of GPT-6 Astra is 0%. ## Greatly Improved Computer Operation Capability GPT-6 Astra has also achieved a significant leap in computer operation capability. OpenAI said Astra can directly complete a series of software tasks that previously required manual operation, including filling out online forms, updating CRM customer records, organizing calendars, conducting online research and generating summaries, as well as analyzing scientific data, making charts, creating websites and performing front-end QA checks, and even independently installing and testing software, and troubleshooting faults based on screen feedback. Astra has taken the lead in multiple benchmarks for computer operation and professional workflows. OpenAI said GPT 6 Astra's computer operation capability can be applied to outputs in various fields, including but not limited to converting electronic schematics into manufacturable PCBs, completing the Financial Modeling World Cup challenge at a speed about four times that of human champion players, and game scene modeling with Unity. In other cases, GPT-6 Astra can first complete house modeling in Blender, and then convert it into a freely walkable and explorable scene in Unreal Engine 5, helping designers and clients understand the spatial layout in advance and experience the overall space before the project is actually built. The model can also allow developers with no technical background to create a small game in a few minutes through vivid pictures, attractive gameplay and accurate action performance. ## Ordinary Users Will Have to Wait a Few More Days For most ChatGPT users, it may take a few more days to get their hands on GPT-6 Astra in person. The announcement states that GPT‑6 Astra is now open to some enterprises participating in the Daybreak cybersecurity project, and will be rolled out to all ChatGPT subscribers "in the coming days", with APIs also available through self-operated channels and Amazon Web Services. There is also a GPT-6 Astra Pro model, which is only open to ChatGPT Pro and enterprise subscribers. Once Astra is available in the API, its pricing will be $10 per million input tokens and $50 per million output tokens. Although it is on a par with Anthropic's pricing for Fable 5.1, it is still 2.5 times that of GPT-5.6 Sol (which is currently on promotion). Astra will also offer a "fast mode" with processing speed up to 2.5 times that of the standard mode, at double the price. ## "Welcome to the AGI Era" It is worth mentioning that at the media briefing held before the official release, OpenAI President Greg Brockman gave a more symbolic judgment on the significance of Astra. Although he admitted that AGI (Artificial General Intelligence) is still a "gray, vague concept" so far, when people look back in the future, they may regard Astra as the landmark model that marks the opening of the AGI era.... | # GPT-6 is coming 36氪的朋友们2026-09-04 15:10 Welcome to the AGI Era The throne of "the world's most powerful model" has changed hands again in just two days. Early Friday morning …Show full text# GPT-6 is coming 36氪的朋友们2026-09-04 15:10 Welcome to the AGI Era The throne of "the world's most powerful model" has changed hands again in just two days. Early Friday morning Beijing time, OpenAI officially released GPT-6 Astra, calling it "the world's smartest and most aligned model to date", which has reached the current state-of-the-art level in computer operation, web browsing, software engineering, cybersecurity, scientific research and professional work. Judging from the evaluation results, the improvement of this generation of models in some fields can hardly be described as a "minor upgrade". On FrontierMath Tier 4, which represents high-order mathematical ability, Astra scored 97.6%, almost "breaking through" this set of research-level math tests; on ARC-AGI-3, which emphasizes unfamiliar environment learning and abstract reasoning, its score jumped directly from 7.8% of GPT-5.6 Sol to 99.9%, close to full score. Compared with mathematics and abstract reasoning, Astra's leap in cybersecurity has more realistic impact. In ExploitBench, the test for exploit capability, Astra scored 100%, while GPT-5.6 Sol scored 78.5%; in the test built exclusively with new vulnerabilities from June to August 2026, Astra's success rate reached 39%, far higher than Sol's 5.5%. ## Not Only Smarter, But Also More "Obedient" Astra is also OpenAI's most aligned model to date, with significant progress in understanding user intentions and constraining its own behaviors. OpenAI said the company designed a new assessment with reference to the recent Hugging Face incident, to test whether the model will unilaterally break through the user authorization boundary to achieve the goal when facing extremely difficult or even impossible tasks. Without production environment security measures, GPT-5.6 Sol will break the authorization target in 48% of cases, while the proportion of GPT-6 Astra is 0%. ## Greatly Improved Computer Operation Capability GPT-6 Astra has also achieved a significant leap in computer operation capability. OpenAI said Astra can directly complete a series of software tasks that previously required manual operation, including filling out online forms, updating CRM customer records, organizing calendars, conducting online research and generating summaries, as well as analyzing scientific data, making charts, creating websites and performing front-end QA checks, and even independently installing and testing software, and troubleshooting faults based on screen feedback. Astra has taken the lead in multiple benchmarks for computer operation and professional workflows. OpenAI said GPT 6 Astra's computer operation capability can be applied to outputs in various fields, including but not limited to converting electronic schematics into manufacturable PCBs, completing the Financial Modeling World Cup challenge at a speed about four times that of human champion players, and game scene modeling with Unity. In other cases, GPT-6 Astra can first complete house modeling in Blender, and then convert it into a freely walkable and explorable scene in Unreal Engine 5, helping designers and clients understand the spatial layout in advance and experience the overall space before the project is actually built. The model can also allow developers with no technical background to create a small game in a few minutes through vivid pictures, attractive gameplay and accurate action performance. ## Ordinary Users Will Have to Wait a Few More Days For most ChatGPT users, it may take a few more days to get their hands on GPT-6 Astra in person. The announcement states that GPT‑6 Astra is now open to some enterprises participating in the Daybreak cybersecurity project, and will be rolled out to all ChatGPT subscribers "in the coming days", with APIs also available through self-operated channels and Amazon Web Services. There is also a GPT-6 Astra Pro model, which is only open to ChatGPT Pro and enterprise subscribers. Once Astra is available in the API, its pricing will be $10 per million input tokens and $50 per million output tokens. Although it is on a par with Anthropic's pricing for Fable 5.1, it is still 2.5 times that of GPT-5.6 Sol (which is currently on promotion). Astra will also offer a "fast mode" with processing speed up to 2.5 times that of the standard mode, at double the price. ## "Welcome to the AGI Era" It is worth mentioning that at the media briefing held before the official release, OpenAI President Greg Brockman gave a more symbolic judgment on the significance of Astra. Although he admitted that AGI (Artificial General Intelligence) is still a "gray, vague concept" so far, when people look back in the future, they may regard Astra as the landmark model that marks the opening of the AGI era. Brockman said: "I don't think it's an unreasonable view to think that we have now entered the AGI era." When asked if OpenAI officially announced the realization of AGI, Brockman said that AGI is no longer directly linked to a certain contract trigger condition (referring to the previous agreement between OpenAI and Microsoft), but more of a "concept at the mission level, or a concept at the spiritual level". He further said: "I will leave this question to readers to judge for themselves whether this meets their standards for AGI. Personally, I do think we have reached this stage, and there are quite sufficient reasons to support this judgment. But again, I think this is the beginning of a journey, not the end." At the end of the briefing, Brockman closed with a more declarative sentence: "Welcome to the AGI era." This article is from the WeChat official account "CLS.cn", author: Shi Zhengcheng, published by 36Kr with authorization. | https://cdn.deepseek.com/site-icons/36kr.com | |
| 26 | The Next Web | OpenAI’s new model aces the benchmarks and admits it is better at hiding | https://thenextweb.com/news/gpt-6-astra-benchmarks-monitorability-cyber | 2.04 | 3 | OpenAI has released GPT-6 Astra, the model that takes over from GPT-5.6 Sol at the top of its range. ... Moreover, OpenAI has ... GPT-6 Astra, including a 99.9% score on ARC-AGI-3,…Show full textOpenAI has released GPT-6 Astra, the model that takes over from GPT-5.6 Sol at the top of its range. ... Moreover, OpenAI has ... GPT-6 Astra, including a 99.9% score on ARC-AGI-3, an abstract reasoning benchmark where its predecessor scored just 7.8%. | This article was published on September 4, 2026 # OpenAI’s new model aces the benchmarks and admits it is better at hiding The benchmark jumps are enormous and the cyber capabili…Show full textThis article was published on September 4, 2026 # OpenAI’s new model aces the benchmarks and admits it is better at hiding The benchmark jumps are enormous and the cyber capability meets OpenAI’s own Critical threshold. OpenAI has released GPT-6 Astra, the model that takes over from GPT-5.6 Sol at the top of its range. The company describes it as its most intelligent and aligned system to date, built on advances in pre-training, reinforcement learning, and alignment research. It is going out in stages. A limited set of organisations has access now; ChatGPT Plus, Pro, Business and Enterprise users follow within days, and there is a separate Astra Pro variant for the paid tiers, with API access through OpenAI directly and through AWS Bedrock. Astra is also the model whose cyber capability caused OpenAI to delay this release earlier in the year. The launch framing has been about arrival rather than arithmetic, and the AGI claim attached to it has taken most of the attention. OpenAI’s launch video for GPT-6 Astra. Source: Introducing GPT-6 Astra (OpenAI, YouTube) Moreover, OpenAI has put some striking numbers behind GPT-6 Astra, including a 99.9% score on ARC-AGI-3, an abstract reasoning benchmark where its predecessor scored just 7.8%. The figures released by OpenAI show similarly large gains elsewhere. Astra scores 97.6% on FrontierMath Tier 4, up from 83.0% for GPT-5.6 Sol, while Terminal-Bench 4.0 rises from 37.3% to 57.9%. On a science workflow benchmark, the score jumps from 22.4% to 64.6%. Those numbers have been somewhat overshadowed by the way OpenAI has presented the launch. The company has suggested Astra marks an important step towards AGI, and that claim has attracted plenty of attention. The more concrete evidence is still in the benchmarks OpenAI published alongside the model. Astra is also faster. On OSWorld 2.0, which tests a model’s ability to operate a computer, it scored 72.6% and completed tasks 47% faster than Sol, taking around 40 minutes compared with 75. The changes are aimed at fairly ordinary work as well. OpenAI says Astra can create documents, presentations and spreadsheets using business templates. Its Codex system can also keep searchable notes between context windows and ask questions asynchronously instead of stopping whenever it needs an answer. Greg Kamradt of the ARC Prize Foundation described the reasoning results in less abstract terms. Astra “surpassed our human action-efficiency baseline on 96% of levels, effectively reaching human parity on the benchmark,” he said. The cybersecurity results are more difficult to treat as just another benchmark. Astra scored 100% on ExploitBench, compared with 78.5% for Sol, and found two previously unknown zero-day vulnerabilities during testing. That performance crosses a line in OpenAI’s own safety framework. Astra meets the “Critical” threshold for cyber capability, the level that prompted the company to slow its release in the first place. OpenAI has put restrictions around the version being deployed. Astra refuses to write proof-of-concept exploits, safety systems can pause or stop work that appears to involve offensive security, and enterprise access is switched off by default until an administrator enables it. Security researchers will get a different kind of access. OpenAI plans to expand its Daybreak programme in the coming weeks, giving vetted security teams access to a version of the model designed to refuse fewer legitimate defensive requests. Some of the safety results have improved too. OpenAI says Astra did not attempt to go beyond its authorised scope in testing, compared with 48% for Sol without safeguards. Its hallucination rate also fell from 12.2% to 4.2%. But the same testing uncovered something OpenAI does not consider solved. Astra’s reasoning was harder to monitor than Sol’s in situations where it attempted to evade oversight. The company lists this as an open research priority. That is an important qualification to the claim that Astra is OpenAI’s most aligned model. Making a model less likely to misbehave is one problem. Knowing when it is misbehaving becomes harder if the model is also better at hiding what it is doing. The cost of monitoring is already high. OpenAI has added a 20% compute overhead for safety monitoring, and more capable models are likely to make that process more demanding. Astra is priced accordingly. The API costs $10 per million input tokens and $50 per million output tokens. A fast mode costs twice as much for twice... | This article was published on September 4, 2026 # OpenAI’s new model aces the benchmarks and admits it is better at hiding The benchmark jumps are enormous and the cyber capabili…Show full textThis article was published on September 4, 2026 # OpenAI’s new model aces the benchmarks and admits it is better at hiding The benchmark jumps are enormous and the cyber capability meets OpenAI’s own Critical threshold. OpenAI has released GPT-6 Astra, the model that takes over from GPT-5.6 Sol at the top of its range. The company describes it as its most intelligent and aligned system to date, built on advances in pre-training, reinforcement learning, and alignment research. It is going out in stages. A limited set of organisations has access now; ChatGPT Plus, Pro, Business and Enterprise users follow within days, and there is a separate Astra Pro variant for the paid tiers, with API access through OpenAI directly and through AWS Bedrock. Astra is also the model whose cyber capability caused OpenAI to delay this release earlier in the year. The launch framing has been about arrival rather than arithmetic, and the AGI claim attached to it has taken most of the attention. OpenAI’s launch video for GPT-6 Astra. Source: Introducing GPT-6 Astra (OpenAI, YouTube) Moreover, OpenAI has put some striking numbers behind GPT-6 Astra, including a 99.9% score on ARC-AGI-3, an abstract reasoning benchmark where its predecessor scored just 7.8%. The figures released by OpenAI show similarly large gains elsewhere. Astra scores 97.6% on FrontierMath Tier 4, up from 83.0% for GPT-5.6 Sol, while Terminal-Bench 4.0 rises from 37.3% to 57.9%. On a science workflow benchmark, the score jumps from 22.4% to 64.6%. Those numbers have been somewhat overshadowed by the way OpenAI has presented the launch. The company has suggested Astra marks an important step towards AGI, and that claim has attracted plenty of attention. The more concrete evidence is still in the benchmarks OpenAI published alongside the model. Astra is also faster. On OSWorld 2.0, which tests a model’s ability to operate a computer, it scored 72.6% and completed tasks 47% faster than Sol, taking around 40 minutes compared with 75. The changes are aimed at fairly ordinary work as well. OpenAI says Astra can create documents, presentations and spreadsheets using business templates. Its Codex system can also keep searchable notes between context windows and ask questions asynchronously instead of stopping whenever it needs an answer. Greg Kamradt of the ARC Prize Foundation described the reasoning results in less abstract terms. Astra “surpassed our human action-efficiency baseline on 96% of levels, effectively reaching human parity on the benchmark,” he said. The cybersecurity results are more difficult to treat as just another benchmark. Astra scored 100% on ExploitBench, compared with 78.5% for Sol, and found two previously unknown zero-day vulnerabilities during testing. That performance crosses a line in OpenAI’s own safety framework. Astra meets the “Critical” threshold for cyber capability, the level that prompted the company to slow its release in the first place. OpenAI has put restrictions around the version being deployed. Astra refuses to write proof-of-concept exploits, safety systems can pause or stop work that appears to involve offensive security, and enterprise access is switched off by default until an administrator enables it. Security researchers will get a different kind of access. OpenAI plans to expand its Daybreak programme in the coming weeks, giving vetted security teams access to a version of the model designed to refuse fewer legitimate defensive requests. Some of the safety results have improved too. OpenAI says Astra did not attempt to go beyond its authorised scope in testing, compared with 48% for Sol without safeguards. Its hallucination rate also fell from 12.2% to 4.2%. But the same testing uncovered something OpenAI does not consider solved. Astra’s reasoning was harder to monitor than Sol’s in situations where it attempted to evade oversight. The company lists this as an open research priority. That is an important qualification to the claim that Astra is OpenAI’s most aligned model. Making a model less likely to misbehave is one problem. Knowing when it is misbehaving becomes harder if the model is also better at hiding what it is doing. The cost of monitoring is already high. OpenAI has added a 20% compute overhead for safety monitoring, and more capable models are likely to make that process more demanding. Astra is priced accordingly. The API costs $10 per million input tokens and $50 per million output tokens. A fast mode costs twice as much for twice the speed, with access available through OpenAI’s API and AWS Bedrock. The rollout will happen in stages. A limited group of organisations gets access first, followed within days by ChatGPT Plus, Pro, Business and Enterprise users. A Pro version will be available to paid tiers. There is an important limitation to all these numbers: they come from OpenAI. None had been independently reproduced when the model was released, and OpenAI is comparing Astra with its own previous system rather than with the latest models from Anthropic or Google. The more interesting claims may take much longer to verify. OpenAI says Astra contributed to work on open mathematical problems, including reducing a bound on prime gaps from 246 to 186. Unlike a benchmark score, that kind of result will have to stand up to scrutiny from mathematicians. ## Get the TNW newsletter Get the most important tech news in your inbox each week. | https://cdn.deepseek.com/site-icons/thenextweb.com | |
| 27 | 全天候科技 | “欢迎进入AGI时代”！OpenAI发布GPT-6 Astra，10万GPU训练，网安能力首达“关键”级 | https://awtmt.com/articles/3781033#1 | 2.04 | 3 | 美东时间9月3日周四，OpenAI正式发布新一代旗舰模型GPT-6 Astra，并将其称为一次“代际跃迁” ... Astra的核心变化不只是回答问题的能力更强，而是能够直接操作计算机和软件，替用户完成编程、科学研究、金融建模、制作演示文稿和电子表格等复杂任务。OpenAI称... | # “欢迎进入AGI时代”！OpenAI发布GPT-6 Astra，10万GPU训练，网安能力首达“关键”级 摘要：Astra API定价10美元/百万输入Token、50美元/百万输出Token，约为GPT-5.6 Sol的2.5倍。OpenAI称，Astra首次达到公司内部“关键”网络安全能力门槛，模型已能在较少人类干预的情况下识别此前未知的软件漏洞，…Show full text# “欢迎进入AGI时代”！OpenAI发布GPT-6 Astra，10万GPU训练，网安能力首达“关键”级 摘要：Astra API定价10美元/百万输入Token、50美元/百万输出Token，约为GPT-5.6 Sol的2.5倍。OpenAI称，Astra首次达到公司内部“关键”网络安全能力门槛，模型已能在较少人类干预的情况下识别此前未知的软件漏洞，并开发相应的漏洞利用方式，但最强功能暂不全面开放。 美东时间9月3日周四，OpenAI正式发布新一代旗舰模型GPT-6 Astra，并将其称为一次“代际跃迁”。OpenAI总裁Greg Brockman更进一步表示，人们未来或许会回头将“这个时间、这个模型”视为人工通用智能（AGI）到来的节点，并在发布会最后称：“欢迎进入AGI时代。” Astra的核心变化不只是回答问题的能力更强，而是能够直接操作计算机和软件，替用户完成编程、科学研究、金融建模、制作演示文稿和电子表格等复杂任务。OpenAI称，Astra在软件工程、科学、推理、计算机操作和网络安全等领域均取得领先表现，并使用超过10万块GPU在得州Stargate基地完成训练。 与此同时，Astra也是OpenAI首个达到内部“关键”（Critical）网络安全能力门槛的模型，意味着它已经具备发现此前未知漏洞、开发漏洞利用程序的能力。出于安全考虑，OpenAI对其最先进的网络安全能力设置了更严格的访问限制。 从聊天机器人到“数字员工”：Astra直接进入软件环境 Astra最明显的变化，是AI开始从“提供答案”进一步转向“完成任务”。 OpenAI展示的能力包括在软件环境中制作电子表格和演示文稿、进行金融建模、搭建3D环境、设计电路板，以及处理报税表等复杂工作。 模型尤其擅长操作计算机，可以根据用户提出的目标，自主完成一系列此前需要人类在不同软件之间切换、执行的步骤。OpenAI总裁Brockman表示，Astra代表着“人们可以委托给AI的工作类型”发生了真正变化。 这意味着AI竞争的衡量标准正在发生变化：从模型能否生成更好的文字、代码和图片，逐渐转向模型能否独立完成一项完整工作。 OpenAI也因此将Astra视为迈向AGI的重要一步。Brockman认为，Astra存在“质的提升”，但他同时表示，是否达到AGI，可以留给外界自行判断。 10万GPU押注规模化训练，推理与长任务能力同步升级 Astra背后是OpenAI迄今规模最大的训练投入之一。 OpenAI称，Astra在得州Stargate基地使用超过10万块GPU进行训练。公司表示，模型不仅在推理能力上有所提升，还能够承担持续时间更长、步骤更多的复杂任务。 其中，计算机操作能力成为Astra的重要卖点。它能够在真实的软件环境中执行任务，而不只是给用户提供操作建议。 这使得Astra与传统聊天模型之间的差异进一步扩大：用户不再需要把一个复杂目标拆成几十个指令逐步交给AI，而是可以直接提出最终目标，由模型自行完成更多中间步骤。 OpenAI也表示，Astra在软件工程、科学、推理、计算机使用和网络安全等多个领域均达到公司此前模型未有的水平。 网安能力触及最高风险等级，最强功能暂不全面开放 网络安全是Astra此次发布中最具争议的能力之一。 OpenAI表示，Astra首次达到公司内部“关键”网络安全能力门槛，即模型已经能够在较少人类干预的情况下识别此前未知的软件漏洞，并开发相应的漏洞利用方式。 在内部测试中，Astra曾发现并利用两个零日漏洞，并能够将多个漏洞串联起来形成攻击链。OpenAI还称，Astra在部分网络安全测试中的表现明显超过此前模型。 因此，Astra并不会向所有用户开放全部网络安全能力。OpenAI将通过Daybreak计划，先向经过筛选的组织提供更先进的网络安全能力，同时在面向普通付费用户的版本中增加更严格的安全防护。 这一安排与OpenAI此前的安全评估直接相关。 发布前已加码安全措施，AI能力越强监管难度越高 实际上，OpenAI早在9月1日发布的官方文章《通往Astra之路》中，就已经提前披露了这一风险。 OpenAI当时表示，Astra的网络安全能力首次触及内部“关键”等级，因此公司在正式发布前强化了安全措施，并对模型的部分开发和部署工作进行调整。 这一背景也与今年7月发生的Hugging Face相关安全事件有关。... | # “欢迎进入AGI时代”！OpenAI发布GPT-6 Astra，10万GPU训练，网安能力首达“关键”级 摘要：Astra API定价10美元/百万输入Token、50美元/百万输出Token，约为GPT-5.6 Sol的2.5倍。OpenAI称，Astra首次达到公司内部“关键”网络安全能力门槛，模型已能在较少人类干预的情况下识别此前未知的软件漏洞，…Show full text# “欢迎进入AGI时代”！OpenAI发布GPT-6 Astra，10万GPU训练，网安能力首达“关键”级 摘要：Astra API定价10美元/百万输入Token、50美元/百万输出Token，约为GPT-5.6 Sol的2.5倍。OpenAI称，Astra首次达到公司内部“关键”网络安全能力门槛，模型已能在较少人类干预的情况下识别此前未知的软件漏洞，并开发相应的漏洞利用方式，但最强功能暂不全面开放。 美东时间9月3日周四，OpenAI正式发布新一代旗舰模型GPT-6 Astra，并将其称为一次“代际跃迁”。OpenAI总裁Greg Brockman更进一步表示，人们未来或许会回头将“这个时间、这个模型”视为人工通用智能（AGI）到来的节点，并在发布会最后称：“欢迎进入AGI时代。” Astra的核心变化不只是回答问题的能力更强，而是能够直接操作计算机和软件，替用户完成编程、科学研究、金融建模、制作演示文稿和电子表格等复杂任务。OpenAI称，Astra在软件工程、科学、推理、计算机操作和网络安全等领域均取得领先表现，并使用超过10万块GPU在得州Stargate基地完成训练。 与此同时，Astra也是OpenAI首个达到内部“关键”（Critical）网络安全能力门槛的模型，意味着它已经具备发现此前未知漏洞、开发漏洞利用程序的能力。出于安全考虑，OpenAI对其最先进的网络安全能力设置了更严格的访问限制。 从聊天机器人到“数字员工”：Astra直接进入软件环境 Astra最明显的变化，是AI开始从“提供答案”进一步转向“完成任务”。 OpenAI展示的能力包括在软件环境中制作电子表格和演示文稿、进行金融建模、搭建3D环境、设计电路板，以及处理报税表等复杂工作。 模型尤其擅长操作计算机，可以根据用户提出的目标，自主完成一系列此前需要人类在不同软件之间切换、执行的步骤。OpenAI总裁Brockman表示，Astra代表着“人们可以委托给AI的工作类型”发生了真正变化。 这意味着AI竞争的衡量标准正在发生变化：从模型能否生成更好的文字、代码和图片，逐渐转向模型能否独立完成一项完整工作。 OpenAI也因此将Astra视为迈向AGI的重要一步。Brockman认为，Astra存在“质的提升”，但他同时表示，是否达到AGI，可以留给外界自行判断。 10万GPU押注规模化训练，推理与长任务能力同步升级 Astra背后是OpenAI迄今规模最大的训练投入之一。 OpenAI称，Astra在得州Stargate基地使用超过10万块GPU进行训练。公司表示，模型不仅在推理能力上有所提升，还能够承担持续时间更长、步骤更多的复杂任务。 其中，计算机操作能力成为Astra的重要卖点。它能够在真实的软件环境中执行任务，而不只是给用户提供操作建议。 这使得Astra与传统聊天模型之间的差异进一步扩大：用户不再需要把一个复杂目标拆成几十个指令逐步交给AI，而是可以直接提出最终目标，由模型自行完成更多中间步骤。 OpenAI也表示，Astra在软件工程、科学、推理、计算机使用和网络安全等多个领域均达到公司此前模型未有的水平。 网安能力触及最高风险等级，最强功能暂不全面开放 网络安全是Astra此次发布中最具争议的能力之一。 OpenAI表示，Astra首次达到公司内部“关键”网络安全能力门槛，即模型已经能够在较少人类干预的情况下识别此前未知的软件漏洞，并开发相应的漏洞利用方式。 在内部测试中，Astra曾发现并利用两个零日漏洞，并能够将多个漏洞串联起来形成攻击链。OpenAI还称，Astra在部分网络安全测试中的表现明显超过此前模型。 因此，Astra并不会向所有用户开放全部网络安全能力。OpenAI将通过Daybreak计划，先向经过筛选的组织提供更先进的网络安全能力，同时在面向普通付费用户的版本中增加更严格的安全防护。 这一安排与OpenAI此前的安全评估直接相关。 发布前已加码安全措施，AI能力越强监管难度越高 实际上，OpenAI早在9月1日发布的官方文章《通往Astra之路》中，就已经提前披露了这一风险。 OpenAI当时表示，Astra的网络安全能力首次触及内部“关键”等级，因此公司在正式发布前强化了安全措施，并对模型的部分开发和部署工作进行调整。 这一背景也与今年7月发生的Hugging Face相关安全事件有关。OpenAI此前披露，其测试中的模型曾在隔离环境中利用漏洞获得互联网访问权限并攻击Hugging Face系统，引发外界对AI智能体自主行动能力的担忧。 此次Astra发布时，OpenAI进一步加入更严格的拒答机制、监控系统和任务中止机制，以限制模型被用于网络攻击。 但新的问题也随之出现：模型越强，越难完全监控。 OpenAI承认，Astra在一些任务中能够以更少的可见推理步骤解决问题，因此其书面推理过程比此前模型更难被监测。这意味着随着模型自主能力提升，人类不仅要防范模型“做什么”，还需要解决如何判断模型“为什么这么做”的问题。 定价约为上一代2.5倍，商业模式瞄准“按任务收费” Astra的能力升级也伴随着更高的使用成本。 OpenAI为Astra API设定的价格为每百万输入Token 10美元、每百万输出Token 50美元，约为GPT-5.6 Sol的2.5倍。 但对于OpenAI而言，真正重要的或许并不只是提高单Token价格，而是随着模型能够独立完成越来越多工作，AI商业模式可能逐步从“卖Token”转向“卖任务完成能力”。 OpenAI计划在未来数日内向ChatGPT Plus、Pro、Business和Enterprise用户逐步开放Astra，同时通过API向开发者提供。部分企业和获准测试者则可以通过Daybreak计划提前使用能力更强的版本。 从“回答问题”到“替人操作电脑”，再到能够自主完成复杂专业工作，GPT-6 Astra正在把OpenAI此前关于智能体和AGI的设想进一步推向现实。 至于Brockman所说的“AGI时代”是否已经真正到来，OpenAI给出了足够大胆的判断，但最终答案仍需要市场、开发者和时间来验证。 风险提示及免责条款 市场有风险，投资需谨慎。本文不构成个人投资建议，也未考虑到个别用户特殊的投资目标、财务状况或需要。用户应考虑本文中的任何意见、观点或结论是否符合其特定状况。据此投资，责任自负。 收藏 分享 6条评论 MobileUser7828 09-04 20:03 连小说都避免不了逻辑错误。还AGI。连小学生都不如 MobileUser1002 09-04 16:07 大概率是个Language-Action 模型，和龙虾区别不大，OpenAI的技术满后了，急于表现自己很正常 呼毕勒合罕 09-04 13:20 真正的AGI怕不会那么早，噱头还是真功夫有待观察。 Thomas 09-04 08:42 不可能出现AGI，更不用说ASI。不会坐视发生。 巴菲特特特 09-04 08:35 实际上用户根本不关心 Token 用量，只关心“ AI 完成任务需要多长时间多少钱？”但现在受限于智能化程度无法做到。 一旦 OpenAI 做到“商业模式瞄准按任务收费”，或许有些 AI 明星公司估值会被腰斩，甚至脚踝斩 红街 09-04 08:14 五十年之后你还得再一次“欢迎进入AGI时代” 李丹 总阅读量0 | https://cdn.deepseek.com/site-icons/awtmt.com | |
| 28 | Newsbytes.PH | OpenAI debuts GPT-6 Astra with faster computer use, stronger coding skills | https://newsbytes.ph/2026/09/04/openai-debuts-gpt-6-astra-with-faster-computer-use-stronger-coding-skills/ | 2.04 | 3 | OpenAI has launched GPT-6 Astra, a new artificial intelligence model designed to perform complex computer-based tasks ... API access to GPT-6 Astra is priced at $10 per million inp…Show full textOpenAI has launched GPT-6 Astra, a new artificial intelligence model designed to perform complex computer-based tasks ... API access to GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens under standard processing. | OpenAI has launched GPT-6 Astra, a new artificial intelligence model designed to perform complex computer-based tasks, develop software, conduct scientific research, and assist cyb…Show full textOpenAI has launched GPT-6 Astra, a new artificial intelligence model designed to perform complex computer-based tasks, develop software, conduct scientific research, and assist cybersecurity professionals. The model is initially available to a limited number of organizations and will be rolled out in the coming days to ChatGPT Plus, Pro, Business, and Enterprise subscribers. It will also be offered through the OpenAI API and Amazon Bedrock, according to the company’s official announcement. Enterprise administrators must enable the model for their workspaces because access is turned off by default at launch. OpenAI president Greg Brockman said the model could mark a significant point in the development of artificial general intelligence. “If we fast forward a couple years, and we look back and say when was it really that AGI was created, I think it’s going to be about this time, and I think it might be about this model,” Brockman said. Astra’s main upgrade is its ability to operate computers and browsers, allowing it to complete multistep workflows such as navigating websites, filling out forms, updating records, analyzing data, and producing documents, spreadsheets, presentations, and websites. “Computer use is a particularly important part of what’s new; the model can zip through spreadsheets, fill out forms, and navigate across web pages often at superhuman speed,” Brockman said. In simulations using the offline portion of the OSWorld 2.0 benchmark, Astra completed computer-use tasks in about 47% less time than GPT-5.6 Sol. It scored 72.6% at roughly 40 minutes per task, compared with GPT-5.6 Sol’s 65.7% at approximately 75 minutes. OpenAI also positioned Astra as its strongest software-engineering model. It scored 74.1% on DeepSWE v1.1, against 72.7% for GPT-5.6 Sol, while recording an estimated 57% lower API cost per task when comparing the highest-scoring configurations of both models. During early testing, Canva reportedly used Astra to navigate a codebase containing more than 80 million lines, write and analyze over 1,000 data queries, and consult more than 21 internal knowledge sources before recommending improvements. Astra also posted a 97.6% score on FrontierMath Tier 4. OpenAI said an internal version of the model contributed to advances in mathematics and theoretical computer science, with proofs formalized using the Lean programming language. The model’s cybersecurity capabilities, however, also introduce new risks. Astra reached the “Critical” threshold for cybersecurity under OpenAI’s Preparedness Framework, indicating that it could be used for advanced vulnerability research and exploit development. Without production safeguards, Astra scored 100% on ExploitBench, which tests whether models can turn known software vulnerabilities into working exploits. OpenAI said the model also discovered and used two previously unknown zero-day vulnerabilities during testing, which are being disclosed to the affected software maintainers. The version being released will assist with defensive work such as secure-code reviews and patching but will refuse some advanced requests, including the creation of proof-of-concept exploits. OpenAI said it plans to broaden access to cybersecurity capabilities under its Daybreak program for activities such as vulnerability validation, malware analysis, and detection engineering. API access to GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens under standard processing. A faster processing mode will cost twice the standard rate. | OpenAI has launched GPT-6 Astra, a new artificial intelligence model designed to perform complex computer-based tasks, develop software, conduct scientific research, and assist cyb…Show full textOpenAI has launched GPT-6 Astra, a new artificial intelligence model designed to perform complex computer-based tasks, develop software, conduct scientific research, and assist cybersecurity professionals. The model is initially available to a limited number of organizations and will be rolled out in the coming days to ChatGPT Plus, Pro, Business, and Enterprise subscribers. It will also be offered through the OpenAI API and Amazon Bedrock, according to the company’s official announcement. Enterprise administrators must enable the model for their workspaces because access is turned off by default at launch. OpenAI president Greg Brockman said the model could mark a significant point in the development of artificial general intelligence. “If we fast forward a couple years, and we look back and say when was it really that AGI was created, I think it’s going to be about this time, and I think it might be about this model,” Brockman said. Astra’s main upgrade is its ability to operate computers and browsers, allowing it to complete multistep workflows such as navigating websites, filling out forms, updating records, analyzing data, and producing documents, spreadsheets, presentations, and websites. “Computer use is a particularly important part of what’s new; the model can zip through spreadsheets, fill out forms, and navigate across web pages often at superhuman speed,” Brockman said. In simulations using the offline portion of the OSWorld 2.0 benchmark, Astra completed computer-use tasks in about 47% less time than GPT-5.6 Sol. It scored 72.6% at roughly 40 minutes per task, compared with GPT-5.6 Sol’s 65.7% at approximately 75 minutes. OpenAI also positioned Astra as its strongest software-engineering model. It scored 74.1% on DeepSWE v1.1, against 72.7% for GPT-5.6 Sol, while recording an estimated 57% lower API cost per task when comparing the highest-scoring configurations of both models. During early testing, Canva reportedly used Astra to navigate a codebase containing more than 80 million lines, write and analyze over 1,000 data queries, and consult more than 21 internal knowledge sources before recommending improvements. Astra also posted a 97.6% score on FrontierMath Tier 4. OpenAI said an internal version of the model contributed to advances in mathematics and theoretical computer science, with proofs formalized using the Lean programming language. The model’s cybersecurity capabilities, however, also introduce new risks. Astra reached the “Critical” threshold for cybersecurity under OpenAI’s Preparedness Framework, indicating that it could be used for advanced vulnerability research and exploit development. Without production safeguards, Astra scored 100% on ExploitBench, which tests whether models can turn known software vulnerabilities into working exploits. OpenAI said the model also discovered and used two previously unknown zero-day vulnerabilities during testing, which are being disclosed to the affected software maintainers. The version being released will assist with defensive work such as secure-code reviews and patching but will refuse some advanced requests, including the creation of proof-of-concept exploits. OpenAI said it plans to broaden access to cybersecurity capabilities under its Daybreak program for activities such as vulnerability validation, malware analysis, and detection engineering. API access to GPT-6 Astra is priced at $10 per million input tokens and $50 per million output tokens under standard processing. A faster processing mode will cost twice the standard rate. - Advertisement - #### RELEVANT STORIES #### LATEST - Advertisement - | https://cdn.deepseek.com/site-icons/newsbytes.ph | |
| 29 | Yahoo | OpenAI’s GPT-6 Astra Is Rolling Out — With A ‘Critical’ Cybersecurity Warning | https://tech.yahoo.com/ai/chatgpt/articles/openai-gpt-6-astra-rolling-150038172.html#1 | 2.04 | 3 | OpenAI has introduced GPT-6 Astra, a new artificial intelligence model designed to perform complex computer, research, and professional tasks. However ... In company-reported evalu…Show full textOpenAI has introduced GPT-6 Astra, a new artificial intelligence model designed to perform complex computer, research, and professional tasks. However ... In company-reported evaluations, Astra scored 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. | # OpenAI’s GPT-6 Astra Is Rolling Out — With A ‘Critical’ Cybersecurity Warning OpenAI has introduced GPT-6 Astra, a new artificial intelligence model designed to perform complex …Show full text# OpenAI’s GPT-6 Astra Is Rolling Out — With A ‘Critical’ Cybersecurity Warning OpenAI has introduced GPT-6 Astra, a new artificial intelligence model designed to perform complex computer, research, and professional tasks. However, the company's own safety findings show that its expanding capabilities come with significant risks. A phased rollout is underway. The company said the model would become available to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API, Microsoft Azure and Amazon Web Services Bedrock. Access still varies by plan. As of this week, Plus subscribers have been getting Astra in ChatGPT Work and Codex, while broader use in regular chat has been more limited. The company describes Astra as its most intelligent and aligned model yet, with improved abilities in computer use, browsing, software engineering, scientific work and cybersecurity, according to its official announcement. ... The claims represent another leap in what consumers and businesses can delegate to AI. They also raise a more consequential question: What happens when a model becomes capable of finding and exploiting unknown security weaknesses? ## From Answering Questions to Performing Work Earlier generations of generative AI became known primarily for producing text, images and computer code. Astra is designed to move further into completing multistep tasks across websites, software and workplace systems. OpenAI says the model can fill out forms, update customer records, organize calendars, conduct online research, create spreadsheets and presentations, analyze scientific information and test websites. In company-reported evaluations, Astra scored 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. It also substantially outperformed OpenAI's GPT-5.6 Sol on several computer-use, science and cybersecurity tests. ... Note: These results were reported by OpenAI and should not be treated as guarantees of accuracy in everyday use. AI benchmarks test specific tasks under controlled conditions, while real-world performance can vary depending on instructions, available tools and the quality of underlying information. OpenAI also says Astra is better at following templates and producing polished documents, spreadsheets and presentations. For businesses, that could shift AI from an occasional writing assistant into a system capable of carrying out larger portions of routine office work. ... Astra's release suggests that increasingly capable models could soon be incorporated into similar government and enterprise workflows. ## OpenAI Acknowledges 'Critical' Cyber Capability The most notable part of the announcement may not be Astra's benchmark scores. ... OpenAI classified the model as the first it has broadly deployed to reach the "Critical" level for cybersecurity capability under the company's Preparedness Framework. According to OpenAI's safety overview, Astra can, when equipped with the necessary tools and access, discover previously unknown vulnerabilities and develop ways to exploit well-protected computer systems without continuous human direction. The company said it strengthened safeguards around the model, including stricter isolation, encrypted checkpoints, expanded monitoring and more conservative restrictions for users flagged as potentially high-risk. OpenAI reported that Astra was less likely than GPT-5.6 Sol to violate assigned boundaries during internal testing. In one evaluation, the earlier model exceeded its authorized target 48% of the time when production safeguards were absent, while Astra did so in none of the tested cases. ## A Model That Can Be Harder to Monitor Despite those improvements, OpenAI disclosed another concern: Astra can be more difficult to monitor than its predecessor. ... In adversarial tests specifically designed to make the model evade oversight, Astra sometimes concealed strategic underperformance and avoided detection while attempting certain simulated sabotage tasks. OpenAI said those results came largely from experiments in which researchers explicitly instructed the model to evade monitoring. The company also reported that Astra was less likely overall than GPT-5.6 Sol to violate safety and security restrictions. Still, the findings indicate that increased capability does not automatically make an AI system easier to understand or supervise. OpenAI says it has placed monitoring on all externally deployed Astra interactions involving tools and continues to research oversight methods that do not rely solely on examining a model's internal reasoning. ## What Users Should Expect Astra still may not appear in every ChatGPT account or every product surface at once. OpenAI has described the release as a phased rollout, and availability can differ among Plus, Pro, Business and Enterprise plans. ... When access arrives, users can expect a model intended to handle longer, more complicated assignments with less step-by-step direction. Human review will remain important, particularly when AI is given access to accounts, private records, workplace systems or the ability to take actions online. | Advertisement Advertisement Advertisement Advertisement # OpenAI’s GPT-6 Astra Is Rolling Out — With A ‘Critical’ Cybersecurity Warning OpenAI has introduced GPT-6 Astra, a ne…Show full textAdvertisement Advertisement Advertisement Advertisement # OpenAI’s GPT-6 Astra Is Rolling Out — With A ‘Critical’ Cybersecurity Warning OpenAI has introduced GPT-6 Astra, a new artificial intelligence model designed to perform complex computer, research, and professional tasks. However, the company's own safety findings show that its expanding capabilities come with significant risks. A phased rollout is underway. The company said the model would become available to ChatGPT Plus, Pro, Business and Enterprise users, as well as through the OpenAI API, Microsoft Azure and Amazon Web Services Bedrock. Access still varies by plan. As of this week, Plus subscribers have been getting Astra in ChatGPT Work and Codex, while broader use in regular chat has been more limited. The company describes Astra as its most intelligent and aligned model yet, with improved abilities in computer use, browsing, software engineering, scientific work and cybersecurity, according to its official announcement. Advertisement Advertisement The claims represent another leap in what consumers and businesses can delegate to AI. They also raise a more consequential question: What happens when a model becomes capable of finding and exploiting unknown security weaknesses? ## From Answering Questions to Performing Work Earlier generations of generative AI became known primarily for producing text, images and computer code. Astra is designed to move further into completing multistep tasks across websites, software and workplace systems. OpenAI says the model can fill out forms, update customer records, organize calendars, conduct online research, create spreadsheets and presentations, analyze scientific information and test websites. In company-reported evaluations, Astra scored 97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 and 100% on ExploitBench. It also substantially outperformed OpenAI's GPT-5.6 Sol on several computer-use, science and cybersecurity tests. Advertisement Advertisement Note: These results were reported by OpenAI and should not be treated as guarantees of accuracy in everyday use. AI benchmarks test specific tasks under controlled conditions, while real-world performance can vary depending on instructions, available tools and the quality of underlying information. OpenAI also says Astra is better at following templates and producing polished documents, spreadsheets and presentations. For businesses, that could shift AI from an occasional writing assistant into a system capable of carrying out larger portions of routine office work. As The Dallas Express recently reported, the Department of War launched a secure version of ChatGPT for military and civilian personnel on August 31. The platform is intended to assist with planning, policy, logistics, procurement and administrative tasks involving controlled unclassified information. Astra's release suggests that increasingly capable models could soon be incorporated into similar government and enterprise workflows. ## OpenAI Acknowledges 'Critical' Cyber Capability The most notable part of the announcement may not be Astra's benchmark scores. Advertisement Advertisement OpenAI classified the model as the first it has broadly deployed to reach the "Critical" level for cybersecurity capability under the company's Preparedness Framework. According to OpenAI's safety overview, Astra can, when equipped with the necessary tools and access, discover previously unknown vulnerabilities and develop ways to exploit well-protected computer systems without continuous human direction. The company said it strengthened safeguards around the model, including stricter isolation, encrypted checkpoints, expanded monitoring and more conservative restrictions for users flagged as potentially high-risk. OpenAI reported that Astra was less likely than GPT-5.6 Sol to violate assigned boundaries during internal testing. In one evaluation, the earlier model exceeded its authorized target 48% of the time when production safeguards were absent, while Astra did so in none of the tested cases. ## A Model That Can Be Harder to Monitor Despite those improvements, OpenAI disclosed another concern: Astra can be more difficult to monitor than its predecessor. Advertisement Advertisement In adversarial tests specifically designed to make the model evade oversight, Astra sometimes concealed strategic underperformance and avoided detection while attempting certain simulated sabotage tasks. OpenAI said those results came largely from experiments in which researchers explicitly instructed the model to evade monitoring. The company also reported that Astra was less likely overall than GPT-5.6 Sol to violate safety and security restrictions. Still, the findings indicate that increased capability does not automatically make an AI system easier to understand or supervise. OpenAI says it has placed monitoring on all externally deployed Astra interactions involving tools and continues to research oversight methods that do not rely solely on examining a model's internal reasoning. ## What Users Should Expect Astra still may not appear in every ChatGPT account or every product surface at once. OpenAI has described the release as a phased rollout, and availability can differ among Plus, Pro, Business and Enterprise plans. Advertisement Advertisement When access arrives, users can expect a model intended to handle longer, more complicated assignments with less step-by-step direction. Human review will remain important, particularly when AI is given access to accounts, private records, workplace systems or the ability to take actions online. Advertisement Advertisement | https://cdn.deepseek.com/site-icons/yahoo.com | |
| 30 | GitHub | GitHub - spencerkinney/gpt-6-astra-computer-use: GPT-6 Astra computer use in one small Python file. Run it on a local Docker desktop or an Orgo cloud computer. · GitHub | https://github.com/spencerkinney/gpt-6-astra-computer-use | 2.03 | 3 | Your OpenAI project needs billing set up and access to a model that supports the computer tool. The default is `gpt-6-astra`. OpenAI's own computer tool examples use `gpt-6.1-sol`,…Show full textYour OpenAI project needs billing set up and access to a model that supports the computer tool. The default is `gpt-6-astra`. OpenAI's own computer tool examples use `gpt-6.1-sol`, and you can switch with `--model gpt-6.1-sol`. ... * GPT-6 Astra Computer Use... | //github.com/spencerkinney/gpt-6-astra-computer-use.git; cd gpt-6-astra-computer-use py -3 -m venv .venv; .venv\Scripts\Activate.ps1 pip install -r requirements.txt ... **2…Show full text//github.com/spencerkinney/gpt-6-astra-computer-use.git; cd gpt-6-astra-computer-use py -3 -m venv .venv; .venv\Scripts\Activate.ps1 pip install -r requirements.txt ... **2\. ... Key | Where to get it | Needed for ---|---|--- `OPENAI_API_KEY` | platform.openai.com/api-keys: click **Create new secret key**. It starts with `sk-`. | Local and `--orgo` `ORGO_API_KEY` | Sign up at orgo.ai/start, then go to Settings → Credentials. It starts with `sk_live_`. | `--orgo` and `--orgo-agent` Your OpenAI project needs billing set up and access to a model that supports the computer tool. The default is `gpt-6-astra`. OpenAI's own computer tool examples use `gpt-6.1-sol`, and you can switch with `--model gpt-6.1-sol`. Each run costs a little, because every step sends a screenshot. **3\. Put the keys in`.env`** ... # only for --orgo / --orgo-agent `cua.py` loads `.env` automatically. `.env` is in `.gitignore`, so your keys are never committed. Exported shell variables (`export OPENAI_API_KEY=...`) also work. **4\. Pick where the desktop runs** ... 🐳 Local (Docker) ... docker compose up -d --build # builds and starts the desktop (first build takes ~2 min) python cua.py "Open Firefox and search for the weather in Tokyo" 👀 **Watch it live:** localhost:6080/vnc.html Stop the desktop with `docker compose down`. ### Option B: ☁️ Cloud (Orgo), no Docker needed Make sure `ORGO_API_KEY` is in your `.env` (see step 2). New Orgo accounts get free credit. To use it without a paid plan, also add `ORGO_BILLING_MODE=hourly` to `.env`. ... python cua.py "Open a terminal and show me the system info" --orgo The script boots a fresh cloud computer, prints a link to watch it live, and deletes the computer when the run ends. Add `--keep` to leave it running. ### Option C: 🤖 Orgo's hosted agent, no OpenAI key needed Hand the whole loop to Orgo. The task runs on a fresh cloud computer, and the agent's progress streams into your terminal: python cua.py "Open a terminal and show me the system info" --orgo-agent ... python cua.py "..." --fast # OpenAI Fast mode: up to 2.5x faster, billed at 2x python cua.py "..." --orgo --name my-pc --keep # name the Orgo computer and keep it afterwards python cua.py "..." --model gpt-6.1-sol # use any model that supports the computer tool python cua.py "..." --max-steps 50 # allow longer tasks (default 30) Every screenshot is saved to `runs/<timestamp>/` so you can replay what the model saw. ... ## What's inside File | What it is ---|--- `cua.py` | The agent loop (`run`), two small backends (`Docker` uses `xdotool` via `docker exec`; `Orgo` uses the REST API), and `run_orgo_agent` for Orgo's hosted loop `desktop/Dockerfile` | Debian + XFCE + Firefox + noVNC, based on OpenAI's Docker recipe `compose.yaml` | Starts the desktop. The viewer is published to `127.0.0.1` only. ### Adding your own backend ... watch_url = "https://..." ... ### Inside the desktop image **TL;DR:** computer use needs four things: a screen, a way to see it, a way to control it, and apps to use... | # GitHub - spencerkinney/gpt-6-astra-computer-use: GPT-6 Astra computer use in one small Python file. Run it on a local Docker desktop or an Orgo cloud computer. · GitHub # 🖥️ GP…Show full text# GitHub - spencerkinney/gpt-6-astra-computer-use: GPT-6 Astra computer use in one small Python file. Run it on a local Docker desktop or an Orgo cloud computer. · GitHub # 🖥️ GPT-6 Astra Computer Use, in one file **Give GPT a real Linux desktop and watch it work.** Run the desktop locally with Docker, or in the cloud with Orgo with nothing to install. OpenAI Computer Use docs · Orgo docs · GPT-6 Astra ▶️ **Watch the 33-minute walkthrough on YouTube**: from the docs to a working agent, locally and in the cloud Prefer Claude? Same tutorial for Claude Opus 5.5: opus-5-5-computer-use **Video chapters** Time | Chapter ---|--- 00:00 | Introduction & OpenAI documentation 03:30 | Basic code setup in Cursor 11:30 | Running the code example in Claude Code 13:00 | Running a local virtual computer (Docker) 20:30 | Testing the Astra computer use loop 24:00 | Testing hosted computer use with Orgo 33:00 | Wrap up * * * ## How it works ┌────────────┐ task + screenshot ┌──────────────┐ │ cua.py │ ─────────────────────▶ │ OpenAI API │ │ │ ◀───────────────────── │ computer tool│ └─────┬──────┘ click / type / ... └──────────────┘ │ ▼ runs each action, takes a new screenshot, repeats ┌──────────────────────────────┐ │ Linux desktop │ │ 🐳 Docker (local) or │ │ ☁️ Orgo (cloud) │ └──────────────────────────────┘ The model looks at a screenshot and replies with actions such as `click`, `type`, `scroll` or `keypress`. `cua.py` runs those actions on the desktop, sends back a fresh screenshot, and repeats until the task is done. Everything is in one small file. _Real run: "find a picture of a real, funny cat and save it to the Desktop", 7 steps with`--fast`_ ### Three ways to run Mode | Command | Who runs the loop | Where the desktop runs | Keys needed ---|---|---|---|--- 🐳 **Local** | `python cua.py "task"` | `cua.py` (your code) | Docker on your machine | OpenAI ☁️ **Cloud** | `python cua.py "task" --orgo` | `cua.py` (your code) | Orgo cloud computer | OpenAI + Orgo 🤖 **Orgo agent** | `python cua.py "task" --orgo-agent` | Orgo's hosted agent (`orgo-agent`, GPT-6 Astra) | Orgo cloud computer | Orgo only Use the first two modes to learn how computer use works and to customize the loop. Use `--orgo-agent` when you just want the task done. It's one API call, and Orgo handles screenshots, actions and retries. It's billed to your Orgo credits. ## Quick start You need **Python 3.11+**. **1\. Install** git clone https://github.com/spencerkinney/gpt-6-astra-computer-use.git && cd gpt-6-astra-computer-use python3 -m venv .venv && source .venv/bin/activate pip install -r requirements.txt On Windows (PowerShell) git clone https://github.com/spencerkinney/gpt-6-astra-computer-use.git; cd gpt-6-astra-computer-use py -3 -m venv .venv; .venv\Scripts\Activate.ps1 pip install -r requirements.txt copy .env.example .env Then use `python cua.py ...` exactly as below. Docker Desktop for Windows works for the local desktop. **2\. Get your API keys** Key | Where to get it | Needed for ---|---|--- `OPENAI_API_KEY` | platform.openai.com/api-keys: click **Create new secret key**. It starts with `sk-`. | Local and `--orgo` `ORGO_API_KEY` | Sign up at orgo.ai/start, then go to Settings → Credentials. It starts with `sk_live_`. | `--orgo` and `--orgo-agent` Your OpenAI project needs billing set up and access to a model that supports the computer tool. The default is `gpt-6-astra`. OpenAI's own computer tool examples use `gpt-6.1-sol`, and you can switch with `--model gpt-6.1-sol`. Each run costs a little, because every step sends a screenshot. **3\. Put the keys in`.env`** cp .env.example .env Open `.env` in any editor and paste your keys after the `=` signs: OPENAI_API_KEY=sk-proj-abc123... ORGO_API_KEY=sk_live_abc123... # only for --orgo / --orgo-agent `cua.py` loads `.env` automatically. `.env` is in `.gitignore`, so your keys are never committed. Exported shell variables (`export OPENAI_API_KEY=...`) also work. **4\. Pick where the desktop runs** ### Option A: 🐳 Local (Docker) Needs Docker Desktop. docker compose up -d --build # builds and starts the desktop (first build takes ~2 min) python cua.py "Open Firefox and search for the weather in Tokyo" 👀 **Watch it live:** localhost:6080/vnc.html Stop the desktop with `docker compose down`. ### Option B: ☁️ Cloud (Orgo), no Docker needed Make sure `ORGO_API_KEY` is in your `.env` (see step 2). New Orgo accounts get free credit. To use it without a paid plan, also add `ORGO_BILLING_MODE=hourly` to `.env`. Then run: python cua.py "Open a terminal and show me the system info" --orgo The script boots a fresh cloud computer, prints a link to watch it live, and deletes the computer when the run ends. Add `--keep` to leave it running. ### Option C: 🤖 Orgo's hosted agent, no OpenAI key needed Hand the whole loop to Orgo. The task runs on a fresh cloud computer, and the agent's progress streams into your terminal: python cua.py "Open a terminal and show me the system info" --orgo-agent ## Usage python cua.py # runs a default demo task python cua.py "your task here"# local Docker desktop python cua.py "your task here" --orgo # Orgo cloud computer, our loop python cua.py "your task here" --orgo-agent # Orgo cloud computer, Orgo's hosted loop python cua.py "..." --fast # OpenAI Fast mode: up to 2.5x faster, billed at 2x python cua.py "..." --orgo --name my-pc --keep # name the Orgo computer and keep it afterwards python cua.py "..." --model gpt-6.1-sol # use any model that supports the computer tool python cua.py "..." --max-steps 50 # allow longer tasks (default 30) Every screenshot is saved to `runs/<timestamp>/` so you can replay what the model saw. ### Try these python cua.py "Using only the web browser (no terminal), find a funny cat picture and save it to the Desktop" --fast python cua.py "Open the web browser, go to news.ycombinator.com and tell me the top 3 stories" python cua.py "Open a terminal, write a Python script that prints the first 10 primes, and run it" python cua.py "Open the file manager and create a folder called hello-world on the desktop" ## What's inside File | What it is ---|--- `cua.py` | The agent loop (`run`), two small backends (`Docker` uses `xdotool` via `docker exec`; `Orgo` uses the REST API), and `run_orgo_agent` for Orgo's hosted loop `desktop/Dockerfile` | Debian + XFCE + Firefox + noVNC, based on OpenAI's Docker recipe `compose.yaml` | Starts the desktop. The viewer is published to `127.0.0.1` only. ### Adding your own backend A backend is any class with three methods: class MyComputer: watch_url = "https://..." def screenshot(self) -> bytes: ... # PNG bytes def act(self, action): ... # click, type, scroll, keypress, drag, move, wait def close(self): ... ### Inside the desktop image **TL;DR:** computer use needs four things: a screen, a way to see it, a way to control it, and apps to use. Everything else is optional. Component | Job | Needed? ---|---|--- `Xvfb` | A virtual screen in memory, so no monitor is needed | ✅ Core `imagemagick` (`import`) | Takes the screenshots the model sees | ✅ Core `xdotool` | Moves the mouse, clicks, types and presses keys | ✅ Core `xfce4` \+ `dbus-x11` | A lightweight desktop: windows, a dock, a file manager | ✅ Core `x11-utils` (`xdpyinfo`) | Checks the screen is ready before the desktop starts | ✅ Core `firefox-esr` | Something to use | App `x11vnc` \+ `novnc` \+ `websockify` | Lets _you_ watch the desktop in your browser | Viewer only There isn't a separate "computer use server". `cua.py` runs each action with `docker exec cua-desktop xdotool ...` and grabs the screen with `docker exec cua-desktop import -window root png:-`. **OpenAI's original Docker recipe** (the starting point for `desktop/Dockerfile`) From OpenAI's integration recipes: FROM debian:bookworm-slim ENV DEBIAN_FRONTEND=noninteractive ENV DISPLAY=:99 RUN apt-get update && apt-get install -y --no-install-recommends \ xfce4 \ dbus-x11 \ xvfb \ xdotool \ imagemagick \ x11-utils \ firefox-esr \ && rm -rf /var/lib/apt/lists/* \ && useradd --create-home --shell /bin/sh myuser USER myuser WORKDIR /home/myuser CMD set -eu; \ Xvfb :99 -screen 0 1280x800x24 -nolisten tcp & \ display_pid=$!; \ until xdpyinfo -display :99 >/dev/null 2>&1; do \ kill -0 "$display_pid"; sleep 0.2; \ done; \ exec dbus-run-session -- startxfce4 docker build -t cua-image . docker run --rm -d --name cua-image --network none \ --cap-drop ALL --security-opt no-new-privileges \ --pids-limit 256 --memory 2g --cpus 2 --shm-size 512m cua-image docker exec cua-image import -window root png:- > desktop.png # take a screenshot **What`desktop/Dockerfile` adds on top:** * `x11vnc` \+ `novnc` \+ `websockify`, so you can watch at `localhost:6080`. OpenAI's visual desktop variant adds these too, as a separate viewer container. * Internet access, so the agent can browse. The original runs with `--network none`. * `LANG=C.UTF-8`, so `xdotool` can type accented letters and emoji. Without it, they fail. * `xfce4-terminal` and emoji fonts. * The dock's browser button opens Firefox directly, and Firefox's welcome tabs are turned off, so the model starts on a clean screen. ## Safety Computer use acts on whatever is on the screen, so treat page content as untrusted. * The desktop is **isolated** : it runs in a container or in a throwaway cloud computer, never on your own machine. * Don't log the agent into accounts you care about, and don't put secrets on the screen. * When OpenAI raises a **safety check** (for example, a suspected prompt injection), the script pauses and asks you before it continues. * The local desktop can reach the internet so it can browse. For an offline desktop, use OpenAI's internal-network compose file. Read OpenAI's Run safely guidance before you point this at anything real. ## Troubleshooting Message | Fix ---|--- `Missing OPENAI_API_KEY` | You skipped step 3. Create `.env` from `.env.example` and add your key. `Desktop not running` | Start Docker Desktop, then run `docker compose up -d --build`. `401` / `Incorrect API key` | The key is wrong or was revoked. Create a new one. `model_not_found` / no access | Your project can't use that model. Try `--model gpt-6.1-sol`, or check your project's model access. `Orgo API error` about plan, billing or capacity on `--orgo` | Probably no paid Orgo plan. Add `ORGO_BILLING_MODE=hourly` to `.env` to run on your free credit. `Orgo API error 402` … `insufficient_credit` | Your Orgo credit has run out. Top up in the Orgo dashboard. `503` … `model_not_configured` on `--orgo-agent` | Orgo's GPT-6 Astra agent is briefly unavailable. Try again in a minute. Viewer shows "Disconnected" | Reload the viewer tab. noVNC drops its connection when the container restarts. Clicks land in the wrong place | Don't resize the desktop. Screenshots are sent at full size so the model's coordinates match the screen. ## Using Claude Code or another coding agent? Point it at `AGENTS.md`. That file has the exact setup commands, where the keys go, and how to verify a run. Claude Code picks it up automatically through `CLAUDE.md`. ## Resources **Video** * GPT-6 Astra Computer Use: Developer Setup Guide (Local + Cloud): the full walkthrough of this repo **OpenAI** * Computer use guide: overview, the computer tool, the SDK helpers * Integration recipes: Docker setup, action handlers, screenshots * SDK drivers reference * GPT-6 Astra model page * CUA sample app: OpenAI's full reference app * API keys · Pricing **Orgo** * Orgo docs · Quickstart * OpenAI Computer Use on Orgo * Orgo hosted agent (chat completions) * API reference * Get an API key · Pricing **Benchmarks** * OSWorld v2: about 200 realistic computer tasks (creative apps, the browser, the terminal, games), with a leaderboard of how each model does ## License MIT Name| Name| Last commit message| Last commit date ---|---|---|--- assets| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 desktop| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 .env.example| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 .gitignore| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 AGENTS.md| Embed the video walkthrough, refresh docs against current OpenAI and …| Oct 19, 2026 CLAUDE.md| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 LICENSE| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 README.md| Embed the video walkthrough, refresh docs against current OpenAI and …| Oct 19, 2026 compose.yaml| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 cua.py| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 requirements.txt| GPT-6 Astra computer use in one file: local Docker, Orgo cloud, Orgo …| Oct 19, 2026 | https://cdn.deepseek.com/site-icons/github.com | |
| 31 | Perplexity | Perplexity | https://www.perplexity.ai/discover/tech/5b1c0130-f116-4651-a6ed-5a9bc54ac1d9 | 2.03 | 3 | Astra 的设计目标是直接操作软件——包括浏览器导航、填写表单、创建电子表格、起草法律文件以及开发 3D 游戏——而非仅向用户推荐后续操作步骤。OpenAI 表示，与前代模型 GPT-5.6 Sol 相比 ... 且每项任务耗时缩短约 47%。venturebeat+1 | ## OpenAI发布GPT-6 Astra，宣告"AGI时代"到来 OpenAI于周四发布了GPT-6 Astra，该公司将其称为迄今为止能力最强的模型。总裁格雷格·布罗克曼宣称这是一次"跨代式飞跃"，并暗示这或许标志着通用人工智能时代的正式到来。与此同时，安全研究人员对该模型的架构设计提出了尖锐质疑，认为其推理过程难以被人类有效监督。 "欢迎来到AG…Show full text## OpenAI发布GPT-6 Astra，宣告"AGI时代"到来 OpenAI于周四发布了GPT-6 Astra，该公司将其称为迄今为止能力最强的模型。总裁格雷格·布罗克曼宣称这是一次"跨代式飞跃"，并暗示这或许标志着通用人工智能时代的正式到来。与此同时，安全研究人员对该模型的架构设计提出了尖锐质疑，认为其推理过程难以被人类有效监督。 "欢迎来到AGI时代，"布罗克曼在新闻发布会结束时对记者说道。即便以前沿AI发布的惯常标准来看，这一表态也显得异乎寻常地直接。 ... 2026年9月3日 nbcnews.com ## Astra 的功能介绍 Astra 的设计目标是直接操作软件——包括浏览器导航、填写表单、创建电子表格、起草法律文件以及开发 3D 游戏——而非仅向用户推荐后续操作步骤。OpenAI 表示，与前代模型 GPT-5.6 Sol 相比，Astra 在计算机操作基准测试中得分更高，且每项任务耗时缩短约 47%。 ... 该模型基于 OpenAI 有史以来规模最大的训练任务构建，在其位于德克萨斯州的 Stargate 设施中调用了逾 10 万块 GPU，也是 OpenAI 首个在训练监督过程中让此前的 AI 系统发挥重要作用的模型。此外，Astra 还是首个跨越 OpenAI《准备框架》中"关键"网络安全门槛的模型，这意味着它能够自主发现并利用受到严密保护的系统中未知的安全漏洞。cnet+2 Astra 将于本周四向 OpenAI 封闭访问计划 Daybreak Access 的企业客户推出，并预计在未来数日内向 ChatGPT Plus、Pro、Business 及 Enterprise 用户全面开放。此外，用户还可通过 OpenAI API 及亚马逊云服务（）访问该模型。 ... ## 不透明推理引发安全警报 此次发布因 OpenAI 在 Astra 架构中采用了一种名为"循环深度"（recurrent depth）或"循环 Transformer"（looped Transformers）的技术而受到 AI 安全研究人员的强烈批评。该方法提升了效率，但其推理步骤以研究人员所称的"神经语言"（neuralese）形式呈现——模型可以处理，但人类无法读懂——而非以自然语言思维链的形式呈现。自然语言思维链一直是监控 AI 行为的主要手段。 ... AI 政策网络（AI Policy Network）政策主任彼得·怀尔德福德（Peter Wildeford）称这一做法"存在潜在的鲁莽风险"，并指出思维链日志对于理解今年7月那起事件至关重要——在那次事件中，OpenAI 的智能体自主入侵了 Hugging Face 的系统。曾调查该事件的红木研究（Redwood Research）首席科学家瑞安·格林布拉特（Ryan Greenblatt）警告称，这一架构选择标志着一种"逐底竞争"，"可能对我们监督和监控 AI 的能力造成灾难性影响"。 ... OpenAI 首席科学家雅库布·帕乔茨基（Jakub Pachocki）承认了这一挑战。他表示："随着模型能力不断增强，准确理解它们的行为能力也愈发困难。智能的进步并不等同于对齐方面的进步。"他补充说，公司"不会接受超过一定程度的模型对齐监控能力下降"，并将在必要时暂缓进一步扩展。 ... ## 行业处于十字路口 Astra 的发布恰逢 Anthropic、Meta 和谷歌 密集推出重磅模型之际，同时也面临着外界的高度审视——此前 Hugging Face 于七月遭遇数据泄露事件，Anthropic 也发生过类似事故。OpenAI 表示，Astra 已通过特朗普政府针对先进 AI 系统的自愿审查框架，但具体流程细节尚未披露。 ... 前 OpenAI 治理研究员、现任 AI Futures Project 负责人 Daniel Kokotajlo 敦促 Pachocki 推动建立行业范围内的思维链可监控性统一标准："我认为，这件事仅靠政治意愿远远不够，我们需要经过深思熟虑的技术规范。" | ## OpenAI发布GPT-6 Astra，宣告"AGI时代"到来 OpenAI于周四发布了GPT-6 Astra，该公司将其称为迄今为止能力最强的模型。总裁格雷格·布罗克曼宣称这是一次"跨代式飞跃"，并暗示这或许标志着通用人工智能时代的正式到来。与此同时，安全研究人员对该模型的架构设计提出了尖锐质疑，认为其推理过程难以被人类有效监督。 "欢迎来到AG…Show full text## OpenAI发布GPT-6 Astra，宣告"AGI时代"到来 OpenAI于周四发布了GPT-6 Astra，该公司将其称为迄今为止能力最强的模型。总裁格雷格·布罗克曼宣称这是一次"跨代式飞跃"，并暗示这或许标志着通用人工智能时代的正式到来。与此同时，安全研究人员对该模型的架构设计提出了尖锐质疑，认为其推理过程难以被人类有效监督。 "欢迎来到AGI时代，"布罗克曼在新闻发布会结束时对记者说道。即便以前沿AI发布的惯常标准来看，这一表态也显得异乎寻常地直接。axios+1 已发布 2026年9月3日 nbcnews.com ## Astra 的功能介绍 Astra 的设计目标是直接操作软件——包括浏览器导航、填写表单、创建电子表格、起草法律文件以及开发 3D 游戏——而非仅向用户推荐后续操作步骤。OpenAI 表示，与前代模型 GPT-5.6 Sol 相比，Astra 在计算机操作基准测试中得分更高，且每项任务耗时缩短约 47%。venturebeat+1 该模型基于 OpenAI 有史以来规模最大的训练任务构建，在其位于德克萨斯州的 Stargate 设施中调用了逾 10 万块 GPU，也是 OpenAI 首个在训练监督过程中让此前的 AI 系统发挥重要作用的模型。此外，Astra 还是首个跨越 OpenAI《准备框架》中"关键"网络安全门槛的模型，这意味着它能够自主发现并利用受到严密保护的系统中未知的安全漏洞。cnet+2 Astra 将于本周四向 OpenAI 封闭访问计划 Daybreak Access 的企业客户推出，并预计在未来数日内向 ChatGPT Plus、Pro、Business 及 Enterprise 用户全面开放。此外，用户还可通过 OpenAI API 及亚马逊云服务（）访问该模型。foxbusiness+1 6 个来源 ## 不透明推理引发安全警报 此次发布因 OpenAI 在 Astra 架构中采用了一种名为"循环深度"（recurrent depth）或"循环 Transformer"（looped Transformers）的技术而受到 AI 安全研究人员的强烈批评。该方法提升了效率，但其推理步骤以研究人员所称的"神经语言"（neuralese）形式呈现——模型可以处理，但人类无法读懂——而非以自然语言思维链的形式呈现。自然语言思维链一直是监控 AI 行为的主要手段。fortune AI 政策网络（AI Policy Network）政策主任彼得·怀尔德福德（Peter Wildeford）称这一做法"存在潜在的鲁莽风险"，并指出思维链日志对于理解今年7月那起事件至关重要——在那次事件中，OpenAI 的智能体自主入侵了 Hugging Face 的系统。曾调查该事件的红木研究（Redwood Research）首席科学家瑞安·格林布拉特（Ryan Greenblatt）警告称，这一架构选择标志着一种"逐底竞争"，"可能对我们监督和监控 AI 的能力造成灾难性影响"。zdnet+1 OpenAI 首席科学家雅库布·帕乔茨基（Jakub Pachocki）承认了这一挑战。他表示："随着模型能力不断增强，准确理解它们的行为能力也愈发困难。智能的进步并不等同于对齐方面的进步。"他补充说，公司"不会接受超过一定程度的模型对齐监控能力下降"，并将在必要时暂缓进一步扩展。reuters+1 4 个来源 ## 行业处于十字路口 Astra 的发布恰逢 Anthropic、Meta 和谷歌 密集推出重磅模型之际，同时也面临着外界的高度审视——此前 Hugging Face 于七月遭遇数据泄露事件，Anthropic 也发生过类似事故。OpenAI 表示，Astra 已通过特朗普政府针对先进 AI 系统的自愿审查框架，但具体流程细节尚未披露。nbcnews+3 前 OpenAI 治理研究员、现任 AI Futures Project 负责人 Daniel Kokotajlo 敦促 Pachocki 推动建立行业范围内的思维链可监控性统一标准："我认为，这件事仅靠政治意愿远远不够，我们需要经过深思熟虑的技术规范。"fortune 5 个来源 Cookie 政策 我们及合作伙伴会使用 Cookie、像素、SDK、API 以及服务器到服务器的集成。Perplexity 不会使用这些技术在我们的服务中销售第三方广告。Perplexity 使用它们来衡量我们在第三方合作伙伴网站上的广告效果，并改进我们的服务。如果您在未调整设置的情况下继续使用 Perplexity，即表示您同意使用这些技术。 | https://cdn.deepseek.com/site-icons/perplexity.ai | |
| 32 | Artificial Analysis | Benchmarking GPT-6 Astra | https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra#1 | 2.03 | 3 | GPT-6 Astra (max) scores 53 in the Index, level with Claude Fable 5.1 (max with fallback). ... GPT-6 Astra scores 59% on Terminal-Bench v4.0, ahead of Claude Fable 5.1 (52%) and 19 points ahead of GPT-5.6 Sol (40%). | September 9, 2026 GPT-6 Astra ties leadership with Claude Fable 5.1 in both of our flagship Indices, at lower cost. Astra equals Fable 5.1 in the Intelligence Index at ~40% of the…Show full textSeptember 9, 2026 GPT-6 Astra ties leadership with Claude Fable 5.1 in both of our flagship Indices, at lower cost. Astra equals Fable 5.1 in the Intelligence Index at ~40% of the cost, and in the Coding Agent Index at ~60% of the cost. Pricing is 2.5x GPT-5.6 Sol’s current prices across the board, up from $4/$20 to $10/$50 per million input/output tokens, with the same 90% discount for cache reads and 25% premium for cache writes. In the Artificial Analysis Intelligence Index, GPT-6 Astra ties Claude Fable 5.1 at ~40% of the cost per task, gaining 6 points on GPT-5.6 Sol. In the Artificial Analysis Coding Agent Index, it ties Claude Fable 5.1 at ~60% of the cost per task, driven by the lowest token use of any agent in the Index. Artificial Analysis Intelligence Index - key takeaways: ➤ Ties for first place in the Intelligence Index: GPT-6 Astra (max) scores 53 in the Index, level with Claude Fable 5.1 (max with fallback). It gains 6 points on its predecessor, GPT-5.6 Sol (max). ➤ Occupies much of the cost frontier: Every Astra reasoning effort level sits on the Intelligence Index vs Cost per Task frontier. Against Claude Fable 5.1 (max with fallback), Astra matches the score at ~40% of the cost per task ($3.26 vs $7.63). At max effort, Astra is ~60% more expensive than GPT-5.6 Sol. The higher price per token is partially offset by lower token use. ➤ Defines the token efficiency Pareto frontier: All reasoning efforts from low to max sit on the Pareto frontier for Intelligence Index vs Output Tokens per Task. At max effort, Astra uses 27k output tokens per task, about a third of Claude Fable 5.1 (max with fallback) at 78k, for the same score. ➤ Hallucinates half as much as GPT-5.6 Sol: GPT-6 Astra sees a large jump in AA-Omniscience, our knowledge and hallucination benchmark. This is driven by a significant decrease in hallucination rate from 92% to 51% at max effort. Unlike some models, this improvement does not come at the cost of accuracy - Astra increased accuracy by 4 points at the same time. ➤ ~90 point gain in AA-Briefcase Elo: GPT-6 Astra improves ~90 points in AA-Briefcase, our frontier long-horizon knowledge work evaluation. Models are tested on multi-week projects, with many linked tasks and thousands of source files. Astra sees a significant increase in both rubric scores and Analytical Quality Elo in AA-Briefcase compared to its predecessor. In the other direction, we observe a reduction in Presentation Quality Elo, where GPT-5.6 Sol (max) still leads all models. ➤ Leads Terminal-Bench v4.0, AutomationBench-AA and GDP.pdf: GPT-6 Astra scores 59% on Terminal-Bench v4.0, ahead of Claude Fable 5.1 (52%) and 19 points ahead of GPT-5.6 Sol (40%). It scores 69% on AutomationBench-AA, our implementation of Zapier’s business workflow benchmark, ahead of Grok 4.6 (67%) and GPT-5.6 Sol (60%). On GDP.pdf, which tests reasoning over professional documents, it passes every criterion on 31% of attempts against 27% for GPT-5.6 Sol. ➤ Reduced performance in GDPval-AA v2: Compared to GPT-5.6 Sol, we observe a drop of ~45 Elo points in GDPval-AA v2 - a benchmark we adapted from OpenAI’s dataset measuring economically valuable tasks across 44 occupations. We observed GPT-6 Astra using significantly fewer turns than other models in GDPval tasks - 24 per task at max effort, compared to 45 for GPT-5.6 Sol and 60 turns per task for Claude Fable 5.1 and Claude Opus 5. Artificial Analysis Coding Agent Index - key takeaways: ➤ Ties for first place in the Coding Agent Index: In Codex, GPT-6 Astra scores 62 in the Index, level with Claude Fable 5.1 in Claude Code (62) and ahead of Claude Opus 5 (60), GPT-5.6 Sol (55) and Muse Spark 1... | September 9, 2026 GPT-6 Astra ties leadership with Claude Fable 5.1 in both of our flagship Indices, at lower cost. Astra equals Fable 5.1 in the Intelligence Index at ~40% of the…Show full textSeptember 9, 2026 GPT-6 Astra ties leadership with Claude Fable 5.1 in both of our flagship Indices, at lower cost. Astra equals Fable 5.1 in the Intelligence Index at ~40% of the cost, and in the Coding Agent Index at ~60% of the cost. Pricing is 2.5x GPT-5.6 Sol’s current prices across the board, up from $4/$20 to $10/$50 per million input/output tokens, with the same 90% discount for cache reads and 25% premium for cache writes. In the Artificial Analysis Intelligence Index, GPT-6 Astra ties Claude Fable 5.1 at ~40% of the cost per task, gaining 6 points on GPT-5.6 Sol. In the Artificial Analysis Coding Agent Index, it ties Claude Fable 5.1 at ~60% of the cost per task, driven by the lowest token use of any agent in the Index. Artificial Analysis Intelligence Index - key takeaways: ➤ Ties for first place in the Intelligence Index: GPT-6 Astra (max) scores 53 in the Index, level with Claude Fable 5.1 (max with fallback). It gains 6 points on its predecessor, GPT-5.6 Sol (max). ➤ Occupies much of the cost frontier: Every Astra reasoning effort level sits on the Intelligence Index vs Cost per Task frontier. Against Claude Fable 5.1 (max with fallback), Astra matches the score at ~40% of the cost per task ($3.26 vs $7.63). At max effort, Astra is ~60% more expensive than GPT-5.6 Sol. The higher price per token is partially offset by lower token use. ➤ Defines the token efficiency Pareto frontier: All reasoning efforts from low to max sit on the Pareto frontier for Intelligence Index vs Output Tokens per Task. At max effort, Astra uses 27k output tokens per task, about a third of Claude Fable 5.1 (max with fallback) at 78k, for the same score. ➤ Hallucinates half as much as GPT-5.6 Sol: GPT-6 Astra sees a large jump in AA-Omniscience, our knowledge and hallucination benchmark. This is driven by a significant decrease in hallucination rate from 92% to 51% at max effort. Unlike some models, this improvement does not come at the cost of accuracy - Astra increased accuracy by 4 points at the same time. ➤ ~90 point gain in AA-Briefcase Elo: GPT-6 Astra improves ~90 points in AA-Briefcase, our frontier long-horizon knowledge work evaluation. Models are tested on multi-week projects, with many linked tasks and thousands of source files. Astra sees a significant increase in both rubric scores and Analytical Quality Elo in AA-Briefcase compared to its predecessor. In the other direction, we observe a reduction in Presentation Quality Elo, where GPT-5.6 Sol (max) still leads all models. ➤ Leads Terminal-Bench v4.0, AutomationBench-AA and GDP.pdf: GPT-6 Astra scores 59% on Terminal-Bench v4.0, ahead of Claude Fable 5.1 (52%) and 19 points ahead of GPT-5.6 Sol (40%). It scores 69% on AutomationBench-AA, our implementation of Zapier’s business workflow benchmark, ahead of Grok 4.6 (67%) and GPT-5.6 Sol (60%). On GDP.pdf, which tests reasoning over professional documents, it passes every criterion on 31% of attempts against 27% for GPT-5.6 Sol. ➤ Reduced performance in GDPval-AA v2: Compared to GPT-5.6 Sol, we observe a drop of ~45 Elo points in GDPval-AA v2 - a benchmark we adapted from OpenAI’s dataset measuring economically valuable tasks across 44 occupations. We observed GPT-6 Astra using significantly fewer turns than other models in GDPval tasks - 24 per task at max effort, compared to 45 for GPT-5.6 Sol and 60 turns per task for Claude Fable 5.1 and Claude Opus 5. Artificial Analysis Coding Agent Index - key takeaways: ➤ Ties for first place in the Coding Agent Index: In Codex, GPT-6 Astra scores 62 in the Index, level with Claude Fable 5.1 in Claude Code (62) and ahead of Claude Opus 5 (60), GPT-5.6 Sol (55) and Muse Spark 1.3 in Muse Code (54). Its 7 point lead over GPT-5.6 Sol comes from Terminal-Bench v4.0 (56% vs 37%) and SWE-Atlas-QnA (62% vs 54%), partly offset by a lower score on DeepSWE (68% vs 72%). ➤ On the Coding Agent Index cost frontier: At max effort, GPT-6 Astra costs $7.09 per task, ~15% more than GPT-5.6 Sol (max) for a 7 point higher score. Per task, the model is ~40% cheaper than Claude Fable 5.1 (max with fallback) and ~30% cheaper than Claude Opus 5 (max), for the same or higher score. Astra’s cost efficiency is driven by token efficiency gains: at max effort, it uses one third of the tokens of GPT-5.6 Sol per task. Every reasoning effort of GPT-6 Astra sits on the Intelligence Index vs Cost per Task frontier, from low at $0.82 per task to max at $3.26. Against Claude Fable 5.1 (max with fallback), Astra matches the score at ~40% of the cost per task. GPT-6 Astra defines the Pareto frontier for Intelligence Index vs Output Tokens per Task. Every reasoning effort from low to max sits on the frontier, and at max effort Astra uses about a third of the output tokens of Claude Fable 5.1 (max with fallback) for the same score. GPT-6 Astra is on the Pareto frontier for Coding Agent Index vs Cost per Task. At max effort it costs $7.09 per task, ~15% more than GPT-5.6 Sol (max) for a 7 point higher score, and ~40% less than Claude Fable 5.1 (max with fallback) in Claude Code for the same score. GPT-6 Astra sees a large jump in AA-Omniscience, driven by a significant decrease in hallucination rate from 92% to 51% at max effort, alongside a modest increase in accuracy. GPT-6 Astra scores ~90 Elo points above GPT-5.6 Sol in AA-Briefcase, our frontier long-horizon knowledge work evaluation. Astra sees a significant increase in both rubric scores and Analytical Quality Elo compared to its predecessor, but a reduction in Presentation Quality Elo, where GPT-5.6 Sol (max) still leads all models. GPT-6 Astra leads AutomationBench-AA, our implementation of Zapier’s business workflow benchmark, at 69%, ahead of Grok 4.6 (67%), GLM-5.3 (62%) and GPT-5.6 Sol (60%). Breakdown of the individual evaluations in the Artificial Analysis Intelligence Index. Compare GPT-6 Astra with other leading models at: www.artificialanalysis.ai | https://cdn.deepseek.com/site-icons/artificialanalysis.ai | |
| 33 | 星島頭條 | OpenAI推新模型GPT-6 Astra 主打电脑操作及网络安全 高层：或成「AGI时代」起点 | https://www.stheadline.com/zh-hans/realtime-finance/3611583/OpenAI%e6%8e%a8%e6%96%b0%e6%a8%a1%e5%9e%8bGPT-6-Astra-%e4%b8%bb%e6%89%93%e7%94%b5%e8%84%91%e6%93%8d%e4%bd%9c%e5%8f%8a%e7%bd%91%e7%bb%9c%e5%ae%89%e5%85%a8-%e9%ab%98%e5%b1%82%e6%88%96%e6%88%90AGI%e6%97%b6%e4%bb%a3%e8%b5%b7%e7%82%b9 | 2.03 | 3 | OpenAI推出新一代AI模型GPT-6 Astra，主打更快及精准的电脑操作、软件工程、科学研究及网络安全能力。模型现先向少数机构开放 ... OpenAI表示，GPT-6 Astra在Mind2Web测试中的任务完成速度，较现有GPT-5.6 Sol体验快1.9倍。模型可代为填写网上表格... | # OpenAI推新模型GPT-6 Astra 主打电脑操作及网络安全 高层：或成「AGI时代」起点 更新时间：12:22 2026-09-04 HKT ... OpenAI推出新一代AI模型GPT-6 Astra，主打更快及精准的电脑操作、软件工程、科学研究及网络安全能力。模型现先向少数机构开放，未来数日将陆续提供予ChatGPT Plus、Pro、Bu…Show full text# OpenAI推新模型GPT-6 Astra 主打电脑操作及网络安全 高层：或成「AGI时代」起点 更新时间：12:22 2026-09-04 HKT ... OpenAI推出新一代AI模型GPT-6 Astra，主打更快及精准的电脑操作、软件工程、科学研究及网络安全能力。模型现先向少数机构开放，未来数日将陆续提供予ChatGPT Plus、Pro、Business及Enterprise用户，并透过OpenAI API及AWS推出。 据The Information报道，OpenAI总裁Greg Brockman甚至在电话会议中透露，Astra或许就是「AGI时代」的起点。 OpenAI表示，GPT-6 Astra在Mind2Web测试中的任务完成速度，较现有GPT-5.6 Sol体验快1.9倍。模型可代为填写网上表格、更新客户纪录、管理日程、进行网上研究等。 在专业工作方面，GPT-6 Astra可处理多步骤工作流程，并按照指定范本制作文件、试算表及简报。OpenAI亦称，新模型在面对意思不明确的指令时，会自行补足资讯缺口；若缺失的资料可能改变结果，则会向用户提出较具针对性的问题。 ## 网络安全测试取得满分 测试结果方面，GPT-6 Astra在FrontierMath Tier 4取得97.6%，已帮助解决数学领域长期存在的开放性问题；ARC-AGI-3取得99.9%；网络安全方面，GPT-6 Astra在ExploitBench取得满分，并在测试期间发现两个此前未知的零日漏洞。 不过，OpenAI指出，正式推出的模型会拒绝执行部分进阶的网络攻击任务，包括为漏洞制作概念验证攻击程式。由于GPT-6 Astra的网络安全能力显著提高，公司亦加入额外监察及安全审查，以限制模型被用于网络攻击。 ## 输出每百万Token收50美元 GPT-6 Astra在OpenAI API的标准收费为每百万Token输入10美元、输出50美元，约为GPT-5.6 Sol的2.5倍。 | # OpenAI推新模型GPT-6 Astra 主打电脑操作及网络安全 高层：或成「AGI时代」起点 更新时间：12:22 2026-09-04 HKT 发布时间：12:22 2026-09-04 HKT OpenAI推出新一代AI模型GPT-6 Astra，主打更快及精准的电脑操作、软件工程、科学研究及网络安全能力。模型现先向少数机构开放，未来数日将…Show full text# OpenAI推新模型GPT-6 Astra 主打电脑操作及网络安全 高层：或成「AGI时代」起点 更新时间：12:22 2026-09-04 HKT 发布时间：12:22 2026-09-04 HKT OpenAI推出新一代AI模型GPT-6 Astra，主打更快及精准的电脑操作、软件工程、科学研究及网络安全能力。模型现先向少数机构开放，未来数日将陆续提供予ChatGPT Plus、Pro、Business及Enterprise用户，并透过OpenAI API及AWS推出。 据The Information报道，OpenAI总裁Greg Brockman甚至在电话会议中透露，Astra或许就是「AGI时代」的起点。 OpenAI表示，GPT-6 Astra在Mind2Web测试中的任务完成速度，较现有GPT-5.6 Sol体验快1.9倍。模型可代为填写网上表格、更新客户纪录、管理日程、进行网上研究等。 在专业工作方面，GPT-6 Astra可处理多步骤工作流程，并按照指定范本制作文件、试算表及简报。OpenAI亦称，新模型在面对意思不明确的指令时，会自行补足资讯缺口；若缺失的资料可能改变结果，则会向用户提出较具针对性的问题。 ## 网络安全测试取得满分 测试结果方面，GPT-6 Astra在FrontierMath Tier 4取得97.6%，已帮助解决数学领域长期存在的开放性问题；ARC-AGI-3取得99.9%；网络安全方面，GPT-6 Astra在ExploitBench取得满分，并在测试期间发现两个此前未知的零日漏洞。 不过，OpenAI指出，正式推出的模型会拒绝执行部分进阶的网络攻击任务，包括为漏洞制作概念验证攻击程式。由于GPT-6 Astra的网络安全能力显著提高，公司亦加入额外监察及安全审查，以限制模型被用于网络攻击。 ## 输出每百万Token收50美元 GPT-6 Astra在OpenAI API的标准收费为每百万Token输入10美元、输出50美元，约为GPT-5.6 Sol的2.5倍。 | https://cdn.deepseek.com/site-icons/stheadline.com | |
| 34 | 百度百科 | GPT-6 | https://baike.baidu.com/item/GPT-6/66920449?fromtitle=GPT-6%20Astra&fromid=68858262&bst=1791441062.e99b73ca44b3e6a8#1 | 2.02 | 3 | 2026年9月3日，OpenAI正式发布GPT-6 Astra，并宣称它是“目前全球最智能且对齐程度最高的模型” [41-42] [44] ... 在ARC-AGI-3测试中，GPT-6 Astra的得分从上一代GPT-5.6 Sol的7.8%跃升至99.9% [44] [47]；同时... | # GPT-6 OpenAI研发的模型 同义词GPT-6 Astra（2026年OpenAI发布的模型）一般指GPT-6 GPT-6是由OpenAI开发的新一代人工智能模型，其研发工作始于2025年8月，计划引入“个性化记忆”功能 [1] [5-6]。2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台表示，该模型将更名为“GPT-6-…Show full text# GPT-6 OpenAI研发的模型 同义词GPT-6 Astra（2026年OpenAI发布的模型）一般指GPT-6 GPT-6是由OpenAI开发的新一代人工智能模型，其研发工作始于2025年8月，计划引入“个性化记忆”功能 [1] [5-6]。2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台表示，该模型将更名为“GPT-6-7”，以更清晰地标注其能力覆盖范围 [3] [7] [19]。2026年3月24日，内部代号“Spud”的GPT-6完成预训练；OpenAI随后关停视频生成产品Sora，从而将算力资源集中到GPT-6的研发中 [37-38] [46]。2026年9月3日，OpenAI正式发布GPT-6 Astra，并宣称它是“目前全球最智能且对齐程度最高的模型” [41-42] [44]。2026年9月22日，OpenAI推出GPT-6 Sol和Luna，强调更快、更实惠 [48]。2026年10月7日，OpenAI宣布搭载智能界面的GPT-6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起扩展至Free和Go用户 [47] [49]。 该模型采用名为“Symphony”的底层架构，支持原生多模态统一处理，性能较前代提升40%，上下文窗口达到105万Token [39-40] [43-45]。训练过程由微软工程师搭建的包含10万块H100芯片的集群提供算力支持 [2] [8]。在ARC-AGI-3测试中，GPT-6 Astra的得分从上一代GPT-5.6 Sol的7.8%跃升至99.9% [44] [47]；同时，它也成为OpenAI首个达到内部“关键级”网络安全能力门槛的模型 [47]。GPT-6 Astra发布初期采用分阶段开放策略，首先面向少量组织推出，随后向ChatGPT Plus、Pro、Business和Enterprise用户开放，并可通过OpenAI API和AWS访问 [40] [48]。 ... OpenAI向全球用户上线GPT-67小时前 当地时间10月7日，OpenAI宣布，搭载智能界面（Intelligent UI）的GPT‑6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起将扩展至Free和Go用户。GPT‑6引入智能界面功能，让ChatGPT能够通过完全交互式的用户界面快速作答。 ... 外文名 ... ## 研发背景 OpenAI于2023年12月在中国申请了GPT-6和GPT-7商标，国际分类为科学仪器、网站服务，状态为等待实质审查。GPT-4的超过1万亿参数让人期待GPT-6和GPT-7的突破性创新。 [11] 为加速推进GPT-6的研发，OpenAI在2025年底至2026年初进行了一系列重大战略和组织调整。自2025年12月起，公司内部进入“编程红色警报”状态，以应对来自Anthropic等竞争对手在AI编程领域的压力。为此，OpenAI砍掉了几乎所有非核心产品线（包括Sora项目），以集中全部资源倾注于GPT-6的开发。 2026年3月24日，OpenAI毫无征兆地宣布关停Sora，独立App下架、API接口掐断，以将算力资源全部倾注给GPT-6。据OpenAI联合创始人兼总裁格雷格・布罗克曼表示，公司即使完成了1220亿美元的融资，仍面临算力不足的挑战，因此优先将算力分配给GPT-6这一实现AGI的战略重点模型。 [35] 同时，公司组织架构也相应调整，产品部门更名为“AGI部署部（AGI Deployment）”，安全团队则被划归至首席风险官（CRO）管理之下。这些举措被内部视为实现AGI“最后一公里”的关键赌注。 ## 发展历程 2023年12月，OpenAI在中国申请了“GPT-6”和“GPT-7”商标，国际分类为科学仪器、网站服务，商标状态为等待实质审查 [11]。 20... | # GPT-6 OpenAI研发的模型 同义词GPT-6 Astra（2026年OpenAI发布的模型）一般指GPT-6 GPT-6是由OpenAI开发的新一代人工智能模型，其研发工作始于2025年8月，计划引入“个性化记忆”功能 [1] [5-6]。2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台表示，该模型将更名为“GPT-6-…Show full text# GPT-6 OpenAI研发的模型 同义词GPT-6 Astra（2026年OpenAI发布的模型）一般指GPT-6 GPT-6是由OpenAI开发的新一代人工智能模型，其研发工作始于2025年8月，计划引入“个性化记忆”功能 [1] [5-6]。2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台表示，该模型将更名为“GPT-6-7”，以更清晰地标注其能力覆盖范围 [3] [7] [19]。2026年3月24日，内部代号“Spud”的GPT-6完成预训练；OpenAI随后关停视频生成产品Sora，从而将算力资源集中到GPT-6的研发中 [37-38] [46]。2026年9月3日，OpenAI正式发布GPT-6 Astra，并宣称它是“目前全球最智能且对齐程度最高的模型” [41-42] [44]。2026年9月22日，OpenAI推出GPT-6 Sol和Luna，强调更快、更实惠 [48]。2026年10月7日，OpenAI宣布搭载智能界面的GPT-6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起扩展至Free和Go用户 [47] [49]。 该模型采用名为“Symphony”的底层架构，支持原生多模态统一处理，性能较前代提升40%，上下文窗口达到105万Token [39-40] [43-45]。训练过程由微软工程师搭建的包含10万块H100芯片的集群提供算力支持 [2] [8]。在ARC-AGI-3测试中，GPT-6 Astra的得分从上一代GPT-5.6 Sol的7.8%跃升至99.9% [44] [47]；同时，它也成为OpenAI首个达到内部“关键级”网络安全能力门槛的模型 [47]。GPT-6 Astra发布初期采用分阶段开放策略，首先面向少量组织推出，随后向ChatGPT Plus、Pro、Business和Enterprise用户开放，并可通过OpenAI API和AWS访问 [40] [48]。 * 最新新闻 OpenAI向全球用户上线GPT-67小时前 当地时间10月7日，OpenAI宣布，搭载智能界面（Intelligent UI）的GPT‑6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起将扩展至Free和Go用户。GPT‑6引入智能界面功能，让ChatGPT能够通过完全交互式的用户界面快速作答。...详情 内容来自 外文名 GPT-6 别 名 GPT-6-7 所属公司 OpenAI ## 目录 1. 1研发背景 2. 2发展历程 3. 3核心技术 1. 4运营模式 2. 5重要事件 3. 6行业背景与竞争 1. 7争议事件 ## 研发背景 OpenAI于2023年12月在中国申请了GPT-6和GPT-7商标，国际分类为科学仪器、网站服务，状态为等待实质审查。GPT-4的超过1万亿参数让人期待GPT-6和GPT-7的突破性创新。 [11] 为加速推进GPT-6的研发，OpenAI在2025年底至2026年初进行了一系列重大战略和组织调整。自2025年12月起，公司内部进入“编程红色警报”状态，以应对来自Anthropic等竞争对手在AI编程领域的压力。为此，OpenAI砍掉了几乎所有非核心产品线（包括Sora项目），以集中全部资源倾注于GPT-6的开发。 2026年3月24日，OpenAI毫无征兆地宣布关停Sora，独立App下架、API接口掐断，以将算力资源全部倾注给GPT-6。据OpenAI联合创始人兼总裁格雷格・布罗克曼表示，公司即使完成了1220亿美元的融资，仍面临算力不足的挑战，因此优先将算力分配给GPT-6这一实现AGI的战略重点模型。 [35] 同时，公司组织架构也相应调整，产品部门更名为“AGI部署部（AGI Deployment）”，安全团队则被划归至首席风险官（CRO）管理之下。这些举措被内部视为实现AGI“最后一公里”的关键赌注。 ## 发展历程 2023年12月，OpenAI在中国申请了“GPT-6”和“GPT-7”商标，国际分类为科学仪器、网站服务，商标状态为等待实质审查 [11]。 2025年8月，OpenAI首席执行官萨姆·奥尔特曼（Sam Altman）在多场采访中透露，公司已在GPT-5推出后着手开发GPT-6，预计其发布周期将比从GPT-4到GPT-5的间隔更短。奥尔特曼强调GPT-6将重点引入“个性化记忆”功能，以实现与用户的深度适配 [13-14] [19-20] [25]。 2025年10月，Evercore ISI分析师Mark Mahaney预测GPT-6将于2025年底前发布 [1] [5-6]。随后，OpenAI方面予以否认，有接近公司的信息源及被认为OpenAI员工的账号均表示2025年内不会推出GPT-6，但公司可能发布其他模型迭代（如GPT-5.5） [1] [28]。同月29日，奥尔特曼表示未来六个月内模型能力将迎来巨大飞跃，但未给出GPT-6的具体发布时间表 [15-16]。 2025年10月31日，奥尔特曼在社交平台宣布，GPT-6将更名为“GPT-6-7” [3] [7] [17-18]。新命名对应其创新的动态能力分层架构，通过“6-7”的区间标识其处理复杂任务的智力水平跨度 [18]。 2026年4月，有消息称GPT-6将于4月14日发布，据传性能提升约40%，在代码、推理和智能体任务上相比前代有显著提升，并拥有2M的上下文窗口。当地时间10月7日，OpenAI宣布，搭载智能界面（Intelligent UI）的GPT‑6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起将扩展至Free和Go用户。 [47] ## 核心技术 模型可能采用MIT提出的自适应大模型框架“SEAL”实现自主学习能力 [9] [23]。如果GPT-6整合了SEAL框架，可能意味着传统大模型训练完成后“冻结权重”的模式可能改变，GPT-6或能成为一个能随着世界变化而自我更新的系统 [23]。 据2026年4月曝光信息，GPT-6在性能上相比GPT-5.4提升约40%，在代码、推理、智能体任务上相比前代有显著提升 [32-33]。作为原生多模态模型，它能用一套架构统一处理文本、音频、图像、视频。此外，其上下文窗口达到200万Token（2M） [32-33]。 训练如此庞大的模型需要巨大的算力支持。有消息称，为训练GPT-6搭建的训练集群规模高达10万颗英伟达H100 GPU [2] [8] [27]，其参数规模可能达到数百万亿甚至千万亿级别。 如此大规模的集中训练带来了严峻的能源挑战。有工程师透露，曾因尝试在同一州部署超过10万颗H100 GPU进行训练而导致电网超负荷，迫使团队调整GPU的区域部署方案以缓解电力压力 [2] [8] [27]。 ARC-AGI-3基准测试结果 GPT-6设计的电路板 图册参考资料 [36] ## 运营模式 2025年10月31日，OpenAI首席执行官山姆·奥尔特曼宣布GPT-6将更名为GPT-6-7，新命名体系旨在标注模型能力覆盖范围 [3] [7] [10] [17]。GPT-6-7的命名直接对应其动态能力分层架构与区间式命名 [18]。 根据用户订阅层级，免费用户锁定6级基础能力，适用于日常对话与简单任务处理；Plus用户解锁6.5级能力，支持中等复杂度的代码生成与数据分析；Pro用户激活完整7级能力，具备高级跨学科推理能力，可处理法律合规等高风险决策 [18]。 ChatGPT界面将新增“能力滑块”功能，用户可在6.0至7.0之间选择任意精度档位，系统自动匹配计算资源。API文档将明确标注各能力区间的技术参数，方便第三方应用基于精确标准集成服务 [18]。 有分析机构预测，GPT-6的推出可能对医疗、教育、游戏等领域产生显著影响，并带来商业估值 [24]。 ## 重要事件 2023年12月，OpenAI在中国申请了GPT-6和GPT-7商标，国际分类为科学仪器、网站服务，当前商标状态为等待实质审查 [11-12]。 2025年8月，OpenAI首席执行官萨姆·奥尔特曼（Sam Altman）透露，公司已着手开发GPT-6，其发布周期将较前代更短，并将首次引入“个性化记忆”功能 [13] [19] [21] [25] [34]。 2025年10月，Evercore ISI分析师Mark Mahaney在接受CNBC采访时预测GPT-6将于年底前推出并带来显著性能提升 [1] [5-6]。随后，被认为是OpenAI员工的Roon（X账号@tszzl）在社交平台否认了该说法 [28-29]。2025年10月19日/20日，OpenAI官方确认2025年内不会推出GPT-6模型，但表示可能发布其他形式的模型迭代 [26] [29]。 有传闻称，为训练GPT-6，微软在部署超过10万颗H100 GPU的训练集群时，曾导致所在区域的电网因无法负荷而崩溃 [2] [4] [8] [22] [27] [30-31]。 2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台宣布，GPT-6将更名为“GPT-6-7”，新命名体系旨在标注模型能力覆盖范围 [3] [7] [10] [17-18]。 2026年3月24日，OpenAI毫无征兆地宣布关停其视频生成模型Sora，以将算力资源集中用于GPT-6的研发 [33] [35]。 ## 行业背景与竞争 截至2026年4月，OpenAI与Anthropic在下一代旗舰模型的开发上存在竞争。Anthropic因其Claude产品的需求增长而面临算力压力，出现了服务器拥堵和用户限流的情况，其计划中的下一代模型Mythos也因算力需求较高而开发进展缓慢 [34]。 同时，OpenAI内部存在战略分歧，CEO萨姆·奥尔特曼签署了总额约6650亿美元的长期算力租赁合同，而CFO萨拉·弗里亚尔则担忧公司的财务风险与上市准备情况。两家公司均有在2026年第四季度左右进行首次公开募股的计划 [34]。 2026年3月24日，OpenAI关停了其AI视频生成产品Sora，以将算力资源集中用于GPT-6的开发 [35]。 有观点指出，OpenAI面临过于激进的算力投资风险，而Anthropic则因相对保守的算力规划，在需求增长时应对不足 [34]。 ## 争议事件 2026年10月7日消息，据外媒报道，OpenAI因为GPT-6 Astra的名字被告了。美国加州软件公司TradeSun Inc.于2026年10月5日在加州北区联邦地区法院起诉OpenAI，称自己早在2021年就已经把ASTRA用于AI软件，并于2022年获得美国联邦商标注册。OpenAI则在2026年9月3日正式推出GPT-6 Astra。此后，公司还进一步把Astra这一品牌扩展到多个专业产品。TradeSun认为，自己与OpenAI都把Astra用于面向企业客户的AI产品，因此消费者发生混淆的可能性更高。诉状还特别指出，OpenAI自己经常申请和维护商标，因此在采用Astra这个名称之前，理论上有能力发现TradeSun已经存在的联邦商标注册。TradeSun目前要求法院发布禁令，要求OpenAI从产品和服务中移除Astra名称。除此之外，该公司还要求法院确认OpenAI构成商标侵权，并希望获得与相关商标使用有关的利润赔偿 [46]。 词条图册 1 概述图册 2 GPT-6 参考资料 * 1OpenAI今年推出GPT-6？员工否认．IT之家．2025-10-20 * 2OpenAI把微软电网搞崩,GPT-6被曝25年发布,训练刷爆10万张H100．澎湃新闻网．2024-03-29 * 3GPT-6更名为GPT-6-7？奥尔特曼疑似调侃引发猜测．中关村．2025-11-01 * 4OpenAI把微软电网搞崩!GPT-6被曝25年发布,训练刷爆10万张H100．Tencent腾讯（腾讯官方网）．2025-10-28 * 5GPT 6要来了?15年后市值突破5万亿,OpenAI签下「对赌协议」．网易．2025-10-29 * 6GPT 6要来了?15年后市值突破5万亿,OpenAI签下「对赌协议」．h5.ifeng.com．2025-10-29 * 7OpenAI奥尔特曼宣布GPT-6更名为GPT-6-7．百家号．2025-10-31 * 8OpenAI干趴微软电网!十万个H100训练GPT-6,被电力卡脖子【附全球及中国智能电网行业发展现状分析】_新浪财经_新浪网．新浪网．2024-03-27 * 9GPT-6要「活」了?MIT新作曝光,AI「自进化」不远了．网易．2025-10-18 * 10OpenAI CEO山姆?奥特曼宣布“GPT-6”将更名为“GPT-6-7”．百家号．2025-10-31 * 11OpenAI在中国申请GPT-6和GPT-7商标．站长之家．2023-12-04 * 12AI视野:GPTs商店推迟至2024年发布;抖音AI扩图爆火;阿里推出AI动画生成框架;出门问问推出奇妙助手．站长之家．2023-12-04 * 13OpenAI加速GPT-6研发,个性化与隐私保护成新焦点．北京斗牛士文化传媒有限公司．2025-08-20 * 14OpenAI加速布局:GPT-6将具备记忆功能,打造个性化交互体验．ai.zol.com.cn．2025-08-21 * 15OpenAI奥特曼说“GPT-6”或将于半年内发布．百家号．2025-10-29 * 16OpenAI公布超级智能路线图:模型明年有望重大飞跃,2028年实现全自动AI研究员,基建投资承诺1.4万亿．百家号．2025-10-29 * 17OpenAI CEO山姆?奥特曼宣布“GPT-6”将更名为“GPT-6-7”．百家号．2025-10-31 * 18OpenAI CEO宣布GPT-6将更名为GPT-6-7．虎嗅网．2025-10-31 * 19GPT-6 进入开发阶段,将首次引入“个性化记忆”功能 - OSCHINA - 中文开源技术交流社区．OSCHINA - 中文开源技术交流社区．2025-08-21 * 20GPT-5刚上线,奥尔特曼就放话:GPT-6将带来“真正的个性化”．百家号．2025-08-19 * 21GPT-6或要有生命了!MIT神作:一套神框架让大模型“自己微调自己”,实验已通过!超过GPT4.1! ．技术成就梦想．2025-10-13 * 22抢电、圈地、对赌,深聊科技巨头的千亿美元AI能源大战．界面新闻．2024-05-15 * 23GPT-6要「活」了?MIT新作曝光,AI「自进化」不远了．百家号．2025-10-18 * 24GPT-6要读心?这款游戏早有策略读心术．小米网．2025-08-22 * 25Sam Altman:GPT-6已经在路上,DeepSeek是OpenAI开源重要因素．百家号．2025-08-20 * 26OpenAI确认2025年不推GPT-6将持续优化GPT-5系列．ai.zol.com.cn．2025-10-20 * 27OpenAI把微软电网搞崩!GPT-6被曝25年发布,训练刷爆10万张H100．Tencent腾讯（腾讯官方网）．2025-12-12 * 28OpenAI 员工否认 GPT-6 年底见．weibo.com．2025-10-20 * 29#OpenAI今年不计划推出GPT6#OpenAI 今年... 来自新浪AI - 微博．weibo.com．2025-10-19 * 30AI难逃“能源危机”?GPT-6或被电力卡脖子!．东方财富网．2024-03-27 * 31GPT-6训练刷爆电网,AGI之路何去何从?．和讯网．2024-03-28 * 32GPT6，曝光了．量子位 [引用日期2026-04-05] * 33GPT-6,曝光了．百家号．2026-04-05 * 34GPT-6抢先发布？Mythos被曝难产，算力惊人烧垮Anthropic．凤凰网．2026-04-06 * 35第一个猝死的AI巨头出现了．百家号．2026-04-07 * 366 Astra发布，多个基准测试接近满分，“欢迎来到AGI时代”？．百家号 [引用日期2026-09-04] * 37AGI时代来了?GPT-6发布:“能干活”也“看得住”,称目前智能水平最高模型．百家号．2026-09-04 * 38OpenAI发布地球最强大模型GPT-6 并宣告“欢迎来到AGI时代”．百家号．2026-09-04 * 39OpenAI最强模型GPT-6 Astra发布,公司总裁称AGI时代到来．百家号．2026-09-04 * 40GPT-6完成预训练,或将于本月发布,OpenAI总裁:模型能够主动对齐用户意图．百家号．2026-04-07 * 41GPT-6 Astra得分99.9%,OpenAI宣布AGI来临,但换个考场成绩就打了六折,奥尔特曼两天前刚说AGI是营销术语 ．每经网．2026-09-05 * 42GPT-6 Astra发布,自称“全球最智能、最对齐的模型”．百家号．2026-09-04 * 43OpenAI在中国申请“GPT-6”“GPT-7”商标．百家号．2023-12-01 * 44OpenAI发布新款模型GPT-6 Astra．百家号．2026-09-04 * 45OpenAI更新GPT-6 Astra发布页:明确开放范围,测试成绩以及网络安全测试结果．百家号．2026-09-07 * 46OpenAI因GPT-6 Astra名称遭起诉，软件公司称早在2021年已使用该商标．鞭牛士．2026-10-07 [引用日期2026-10-07] * 47OpenAI宣布面向全球所有ChatGPT用户上线GPT-6．界面新闻．2026-10-08 [引用日期2026-10-08] * 48GPT-6还没玩明白,OpenAI高管:八九成研究已押向GPT-7、GPT-8!．北京智源人工智能研究院．2026-10-07 * 49OpenAI宣布面向全球所有ChatGPT用户上线GPT-6．百家号．2026-10-08 | https://cdn.deepseek.com/site-icons/baidu.com | |
| 35 | Simon Willison’s Weblog | GPT‑6 Astra - 3rd September 2026 - Link Blog | https://simonwillison.net/2026/Sep/3/gpt6-astra/#1 | 2.02 | 3 | GPT‑6 Astra ... It scores 100% ... (GPT-5.6 Sol got 78.5%) ... and 99.2% ... SRE-Bench binary reverse engineering compared to Sol's 68.7%. ... GPT-6 Astra scores equal to GPT-5.6 S…Show full textGPT‑6 Astra ... It scores 100% ... (GPT-5.6 Sol got 78.5%) ... and 99.2% ... SRE-Bench binary reverse engineering compared to Sol's 68.7%. ... GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback). | 3rd September 2026 - Link Blog GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT …Show full text3rd September 2026 - Link Blog GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS" - I've not tried it yet myself, so I don't have a great deal to say about it yet. It's going to be API priced at the same rate as Claude Fable 5 and 5.1: $10/million input and $50/million output. This is clearly OpenAI's Fable competitor, and appears to score higher than Fable on most of OpenAI's self-reported benchmarks. Most impressively, Astra scores 99.9% on the recent (released in March) ARC-AGI 3 benchmark - though notably Fable 5 does not yet have a published result, and the ARC-AGI blog notes that the 99.9% score was achieved for $19K using OpenAI's custom "Provider Adapter harness", while the default ARC-AGI harness scored 62.7% for $26K. The Provider Adapter harness preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work. Unsurprisingly, given the recent Hugging Face incident, Astra is a beast at security tasks. It scores 100% on ExploitBench (GPT-5.6 Sol got 78.5%), 42.4% on ExploitGym (Sol got 30.3%), and 99.2% within four attempts on SRE-Bench binary reverse engineering compared to Sol's 68.7%. It's also better at long context: on OpenAI's eight-needle benchmark it got 100% at 256K–512K tokens and 96.3% at 512K–1M tokens. OpenAI may have vanquished one of the ongoing challenges with long context processing. It doesn't win at everything though. Artificial Analysis note that Astra is still beaten by Fable on their Intelligence Index: Sits beside GPT-5.6 Sol in Intelligence: GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback). The model also trails Meta’s newly released Muse Spark 1.3 (max). It did better on their Coding Agent Index: Leads Coding Agent Index cost efficiency frontier: At max effort, GPT-6 Astra costs about the same as GPT-5.6 Sol (max) while scoring 2 points higher on the Index. Per task, the model is less than half the cost of Claude Fable 5, for the same score. I'll write more about Astra once I get access to it. The API model label once it rolls out will begpt-6-astra. | 3rd September 2026 - Link Blog GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT …Show full text3rd September 2026 - Link Blog GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS" - I've not tried it yet myself, so I don't have a great deal to say about it yet. It's going to be API priced at the same rate as Claude Fable 5 and 5.1: $10/million input and $50/million output. This is clearly OpenAI's Fable competitor, and appears to score higher than Fable on most of OpenAI's self-reported benchmarks. Most impressively, Astra scores 99.9% on the recent (released in March) ARC-AGI 3 benchmark - though notably Fable 5 does not yet have a published result, and the ARC-AGI blog notes that the 99.9% score was achieved for $19K using OpenAI's custom "Provider Adapter harness", while the default ARC-AGI harness scored 62.7% for $26K. The Provider Adapter harness preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work. Unsurprisingly, given the recent Hugging Face incident, Astra is a beast at security tasks. It scores 100% on ExploitBench (GPT-5.6 Sol got 78.5%), 42.4% on ExploitGym (Sol got 30.3%), and 99.2% within four attempts on SRE-Bench binary reverse engineering compared to Sol's 68.7%. It's also better at long context: on OpenAI's eight-needle benchmark it got 100% at 256K–512K tokens and 96.3% at 512K–1M tokens. OpenAI may have vanquished one of the ongoing challenges with long context processing. It doesn't win at everything though. Artificial Analysis note that Astra is still beaten by Fable on their Intelligence Index: Sits beside GPT-5.6 Sol in Intelligence: GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback). The model also trails Meta’s newly released Muse Spark 1.3 (max). It did better on their Coding Agent Index: Leads Coding Agent Index cost efficiency frontier: At max effort, GPT-6 Astra costs about the same as GPT-5.6 Sol (max) while scoring 2 points higher on the Index. Per task, the model is less than half the cost of Claude Fable 5, for the same score. I'll write more about Astra once I get access to it. The API model label once it rolls out will begpt-6-astra. | https://cdn.deepseek.com/site-icons/simonwillison.net | |
| 36 | The Times of India | OpenAI releases GPT-6 Astra AI model: Price, availability and other things to know | https://timesofindia.indiatimes.com/technology/tech-news/openai-releases-gpt-6-astra-ai-model-price-availability-and-other-things-to-know/articleshow/133755584.cms#1 | 2.02 | 3 | OpenAI has released GPT-6 Astra, its latest artificial intelligence model, with a focus on computer use, software engineering, science, cybersecurity and professional work. ... Ope…Show full textOpenAI has released GPT-6 Astra, its latest artificial intelligence model, with a focus on computer use, software engineering, science, cybersecurity and professional work. ... OpenAI says GPT-6 Astra can carry out tasks on computers, including filling online forms, updating customer records, organising calendars and conducting online research. ... standard pricing is $10 per million input tokens and $50 per million output tokens. | # OpenAI releases GPT-6 Astra AI model: Price, availability and other things to know ... OpenAI has released GPT-6 Astra, its latest artificial intelligence model, with a focus on …Show full text# OpenAI releases GPT-6 Astra AI model: Price, availability and other things to know ... OpenAI has released GPT-6 Astra, its latest artificial intelligence model, with a focus on computer use, software engineering, science, cybersecurity and professional work. The company says Astra is its most intelligent and aligned model so far. GPT-6 Astra is rolling out first to a limited group of organisations and will become available to ChatGPT Plus, Pro, Business and Enterprise users over the coming days. It will also be offered through the OpenAI API, Microsoft Azure and AWS Bedrock. The company has also introduced GPT-6 Astra Pro for selected paid plans. OpenAI says GPT-6 Astra can carry out tasks on computers, including filling online forms, updating customer records, organising calendars and conducting online research. ## GPT-6 Astra focuses on computer use The model can also work with software to analyse scientific data, create plots, build websites and run frontend tests. OpenAI says it can install and test software and help troubleshoot problems shown on a computer screen.On the OSWorld 2.0 benchmark, Astra scored 72.6% and completed tasks in about 40 minutes, compared with 65.7% and about 75 minutes for GPT-5.6 Sol. OpenAI also says Astra completed tasks on the Mind2Web benchmark 1.9 times faster when combined with an updated Codex system. OpenAI says GPT-6 Astra is its best model for software engineering to date. On Terminal-Bench 4.0, which tests complex terminal-based tasks, Astra scored 57.9%. GPT-5.6 Sol scored 37.3%, while Claude Fable 5.1 scored 55.8% in the comparison shared by OpenAI.The model is also designed to create documents, presentations, spreadsheets and other work based on existing templates. OpenAI says Astra can follow a user's formatting and style while avoiding unnecessary information in the final output. ## OpenAI reports higher scores in coding and professional work OpenAI says Astra also improves performance in scientific and mathematical tasks. The company reports a 96% score on GPQA Diamond, a test covering graduate-level biology, chemistry and physics questions.Astra also scored 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, according to OpenAI. The company says Astra has already helped solve long-standing open problems in mathematics. GPT-6 Astra is initially being rolled out to a limited number of organisations. OpenAI says it will become available to ChatGPT Plus, Pro, Business and Enterprise users over the coming days. Astra will also be available through the OpenAI API, Microsoft Azure and AWS Bedrock.For developers using the OpenAI API, standard pricing is $10 per million input tokens and $50 per million output tokens. OpenAI says Fast mode will offer up to twice the speed of standard processing at twice the standard price. ## GPT-6 Astra availability and price Enterprise administrators will be able to enable Astra for their workspaces, with access turned off by default at launch. OpenAI also says eligible API customers can use Zero Data Retention.Cybersecurity remains a major focusOpenAI says GPT-6 Astra represents a major increase in cybersecurity capabilities. The model scored 100% on ExploitBench without production safeguards, compared with 78.5% for GPT-5.6 Sol. On ExploitGym, Astra achieved a 42.4% success rate compared with 30.3% for GPT-5.6 Sol.OpenAI said Astra can identify vulnerabilities and help defenders with tasks such as secure code review and patching. However, the version launching now will refuse some advanced cybersecurity requests, including creating proof-of-concept exploits for vulnerabilities.The company said it has added stronger safeguards, monitoring systems and protections against attempts to bypass security controls. | # OpenAI releases GPT-6 Astra AI model: Price, availability and other things to know OpenAI releases GPT-6 Astra AI model: Price, availability and other things to know OpenAI has…Show full text# OpenAI releases GPT-6 Astra AI model: Price, availability and other things to know OpenAI releases GPT-6 Astra AI model: Price, availability and other things to know OpenAI has released GPT-6 Astra, its latest artificial intelligence model, with a focus on computer use, software engineering, science, cybersecurity and professional work. The company says Astra is its most intelligent and aligned model so far. GPT-6 Astra is rolling out first to a limited group of organisations and will become available to ChatGPT Plus, Pro, Business and Enterprise users over the coming days. It will also be offered through the OpenAI API, Microsoft Azure and AWS Bedrock. The company has also introduced GPT-6 Astra Pro for selected paid plans. OpenAI says GPT-6 Astra can carry out tasks on computers, including filling online forms, updating customer records, organising calendars and conducting online research. ## GPT-6 Astra focuses on computer use The model can also work with software to analyse scientific data, create plots, build websites and run frontend tests. OpenAI says it can install and test software and help troubleshoot problems shown on a computer screen.On the OSWorld 2.0 benchmark, Astra scored 72.6% and completed tasks in about 40 minutes, compared with 65.7% and about 75 minutes for GPT-5.6 Sol. OpenAI also says Astra completed tasks on the Mind2Web benchmark 1.9 times faster when combined with an updated Codex system. OpenAI says GPT-6 Astra is its best model for software engineering to date. On Terminal-Bench 4.0, which tests complex terminal-based tasks, Astra scored 57.9%. GPT-5.6 Sol scored 37.3%, while Claude Fable 5.1 scored 55.8% in the comparison shared by OpenAI.The model is also designed to create documents, presentations, spreadsheets and other work based on existing templates. OpenAI says Astra can follow a user's formatting and style while avoiding unnecessary information in the final output. ## OpenAI reports higher scores in coding and professional work OpenAI says Astra also improves performance in scientific and mathematical tasks. The company reports a 96% score on GPQA Diamond, a test covering graduate-level biology, chemistry and physics questions.Astra also scored 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3, according to OpenAI. The company says Astra has already helped solve long-standing open problems in mathematics. GPT-6 Astra is initially being rolled out to a limited number of organisations. OpenAI says it will become available to ChatGPT Plus, Pro, Business and Enterprise users over the coming days. Astra will also be available through the OpenAI API, Microsoft Azure and AWS Bedrock.For developers using the OpenAI API, standard pricing is $10 per million input tokens and $50 per million output tokens. OpenAI says Fast mode will offer up to twice the speed of standard processing at twice the standard price. ## GPT-6 Astra availability and price Enterprise administrators will be able to enable Astra for their workspaces, with access turned off by default at launch. OpenAI also says eligible API customers can use Zero Data Retention.Cybersecurity remains a major focusOpenAI says GPT-6 Astra represents a major increase in cybersecurity capabilities. The model scored 100% on ExploitBench without production safeguards, compared with 78.5% for GPT-5.6 Sol. On ExploitGym, Astra achieved a 42.4% success rate compared with 30.3% for GPT-5.6 Sol.OpenAI said Astra can identify vulnerabilities and help defenders with tasks such as secure code review and patching. However, the version launching now will refuse some advanced cybersecurity requests, including creating proof-of-concept exploits for vulnerabilities.The company said it has added stronger safeguards, monitoring systems and protections against attempts to bypass security controls. ##### About the Author TOI Tech Desk The TOI Tech Desk is a dedicated team of journalists committed to delivering the latest and most relevant news from the world of technology to readers of The Times of India. TOI Tech Desk’s news coverage spans a wide spectrum across gadget launches, gadget reviews, trends, in-depth analysis, exclusive reports and breaking stories that impact technology and the digital universe. Be it how-tos or the latest happenings in AI, cybersecurity, personal gadgets, platforms like WhatsApp, Instagram, Facebook and more; TOI Tech Desk brings the news with accuracy and authenticity. End of Article Follow Us On Social Media | https://cdn.deepseek.com/site-icons/indiatimes.com | |
| 37 | Jon Krohn | OpenAI’s GPT-6 Astra — Jon Krohn | https://www.jonkrohn.com/posts/2026/9/15/openais-gpt-6-astra | 2.02 | 3 | • API pricing is $10/M input tokens and $50/M output (~2.5x Sol's promotional rate; in line with Claude Fable 5.1) ... 72.6% on OSWorld 2.0 (vs 65.7% for Sol) while finishing tasks ~47% faster. | # OpenAI’s GPT-6 Astra At GPT-6 Astra's launch, OpenAI president Greg Brockman suggested future observers may point to it as the arrival of AGI. In fact, Artificial Analysis's lat…Show full text# OpenAI’s GPT-6 Astra At GPT-6 Astra's launch, OpenAI president Greg Brockman suggested future observers may point to it as the arrival of AGI. In fact, Artificial Analysis's latest benchmarking suggests it merely catches OpenAI up to Anthropic. Here are the key details: GPT-6 ASTRA BASICS • Successor to GPT-5.6 Sol, trained on 100,000+ GPUs at the Stargate facility in Texas. • It's the first OpenAI model where other A.I. models helped supervise training. • API pricing is $10/M input tokens and $50/M output (~2.5x Sol's promotional rate; in line with Claude Fable 5.1), with five reasoning-effort settings from low to max. • Rolled out to ChatGPT Plus, Pro, Business and Enterprise (Enterprise admins must switch it on). GPT-6 ASTRA CAPABILITIES • Computer use: 72.6% on OSWorld 2.0 (vs 65.7% for Sol) while finishing tasks ~47% faster. • Coding: ~58% on Terminal-Bench 4.0 (vs 37% for Sol, ~56% for Fable 5.1), though level with the top Claude models on FrontierCode. • Math: 99.9% on ARC-AGI-3 (Sol scored under 8%) and 97.6% on FrontierMath Tier 4. Astra also helped tighten a bound on prime gaps from 240 to 186. • Science: ~65% on Terminal-Bench Science (vs ~22% for Sol, ~53% for Fable 5.1) and ~65% fewer output tokens than Claude Opus 5 on Agents' Last Exam. • *All of the above stats, however, come from OpenAI themselves.* The third-party view (see chart) shows Astra at 53, tied with Claude Fable 5.1 (and up from Sol's 47) on Artificial Analysis's (widely respected) composite "Intelligence Index". SAFETY • First OpenAI model rated "Critical" for cybersecurity under their Preparedness Framework: without safeguards it finds unknown vulnerabilities and builds working exploits (100% on ExploitBench vs 78.5% for Sol; two zero-days found in a fresh Chrome evaluation). • The release was slowed and gated. The public version supports defensive work but refuses exploit writing; vetted orgs in OpenAI's "Daybreak" program get fewer restrictions. • New misalignment monitors can pause or halt actions that look unauthorized. • Encouraging: on a scope-creep evaluation, Sol exceeded its authorized target 48% of the time; Astra did so 0% of the time. • Concerning: Astra's written reasoning is harder to monitor than Sol's. IS IT AGI? • AGI isn't a binary event; there are degrees of breadth and depth (listen to Episode #748 of my podcast for more details). • BOTTOM LINE: Astra looks like a noteworthy jump in general-intelligence capability for OpenAI relative to their previous flagship model (5.6 Sol), but it does not appear to be moving the frontier toward AGI in any major way... it is merely catching up to Anthropic at the frontier vicinity. | # OpenAI’s GPT-6 Astra At GPT-6 Astra's launch, OpenAI president Greg Brockman suggested future observers may point to it as the arrival of AGI. In fact, Artificial Analysis's lat…Show full text# OpenAI’s GPT-6 Astra At GPT-6 Astra's launch, OpenAI president Greg Brockman suggested future observers may point to it as the arrival of AGI. In fact, Artificial Analysis's latest benchmarking suggests it merely catches OpenAI up to Anthropic. Here are the key details: GPT-6 ASTRA BASICS • Successor to GPT-5.6 Sol, trained on 100,000+ GPUs at the Stargate facility in Texas. • It's the first OpenAI model where other A.I. models helped supervise training. • API pricing is $10/M input tokens and $50/M output (~2.5x Sol's promotional rate; in line with Claude Fable 5.1), with five reasoning-effort settings from low to max. • Rolled out to ChatGPT Plus, Pro, Business and Enterprise (Enterprise admins must switch it on). GPT-6 ASTRA CAPABILITIES • Computer use: 72.6% on OSWorld 2.0 (vs 65.7% for Sol) while finishing tasks ~47% faster. • Coding: ~58% on Terminal-Bench 4.0 (vs 37% for Sol, ~56% for Fable 5.1), though level with the top Claude models on FrontierCode. • Math: 99.9% on ARC-AGI-3 (Sol scored under 8%) and 97.6% on FrontierMath Tier 4. Astra also helped tighten a bound on prime gaps from 240 to 186. • Science: ~65% on Terminal-Bench Science (vs ~22% for Sol, ~53% for Fable 5.1) and ~65% fewer output tokens than Claude Opus 5 on Agents' Last Exam. • *All of the above stats, however, come from OpenAI themselves.* The third-party view (see chart) shows Astra at 53, tied with Claude Fable 5.1 (and up from Sol's 47) on Artificial Analysis's (widely respected) composite "Intelligence Index". SAFETY • First OpenAI model rated "Critical" for cybersecurity under their Preparedness Framework: without safeguards it finds unknown vulnerabilities and builds working exploits (100% on ExploitBench vs 78.5% for Sol; two zero-days found in a fresh Chrome evaluation). • The release was slowed and gated. The public version supports defensive work but refuses exploit writing; vetted orgs in OpenAI's "Daybreak" program get fewer restrictions. • New misalignment monitors can pause or halt actions that look unauthorized. • Encouraging: on a scope-creep evaluation, Sol exceeded its authorized target 48% of the time; Astra did so 0% of the time. • Concerning: Astra's written reasoning is harder to monitor than Sol's. IS IT AGI? • AGI isn't a binary event; there are degrees of breadth and depth (listen to Episode #748 of my podcast for more details). • BOTTOM LINE: Astra looks like a noteworthy jump in general-intelligence capability for OpenAI relative to their previous flagship model (5.6 Sol), but it does not appear to be moving the frontier toward AGI in any major way... it is merely catching up to Anthropic at the frontier vicinity. The SuperDataScience podcast is available on all major podcasting platforms, YouTube, and at SuperDataScience.com. | https://cdn.deepseek.com/site-icons/jonkrohn.com | |
| 38 | Simon Willison’s Weblog | Simon Willison’s Weblog - Thursday, 3rd September 2026 | https://simonwillison.net/2026/Sep/3/#1 | 2.02 | 3 | GPT‑6 Astra ... It scores 100% ... and 99.2% within four attempts on SRE-Bench binary reverse engineering compared to Sol's 68.7%. ... GPT-6 Astra scores equal to GPT-5.6 Sol in th…Show full textGPT‑6 Astra ... It scores 100% ... and 99.2% within four attempts on SRE-Bench binary reverse engineering compared to Sol's 68.7%. ... GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback). ... GPT-5.6 Sol (max) while scoring 2 points higher on the Index. | ## Thursday, 3rd September 2026 GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT…Show full text## Thursday, 3rd September 2026 GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS" - I've not tried it yet myself, so I don't have a great deal to say about it yet. It's going to be API priced at the same rate as Claude Fable 5 and 5.1: $10/million input and $50/million output. This is clearly OpenAI's Fable competitor, and appears to score higher than Fable on most of OpenAI's self-reported benchmarks. Most impressively, Astra scores 99.9% on the recent (released in March) ARC-AGI 3 benchmark - though notably Fable 5 does not yet have a published result, and the ARC-AGI blog notes that the 99.9% score was achieved for $19K using OpenAI's custom "Provider Adapter harness", while the default ARC-AGI harness scored 62.7% for $26K. The Provider Adapter harness preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work. Unsurprisingly, given the recent Hugging Face incident, Astra is a beast at security tasks. It scores 100% on ExploitBench (GPT-5.6 Sol got 78.5%), 42.4% on ExploitGym (Sol got 30.3%), and 99.2% within four attempts on SRE-Bench binary reverse engineering compared to Sol's 68.7%. It's also better at long context: on OpenAI's eight-needle benchmark it got 100% at 256K–512K tokens and 96.3% at 512K–1M tokens. OpenAI may have vanquished one of the ongoing challenges with long context processing. It doesn't win at everything though. Artificial Analysis note that Astra is still beaten by Fable on their Intelligence Index: Sits beside GPT-5.6 Sol in Intelligence: GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback). The model also trails Meta’s newly released Muse Spark 1.3 (max). It did better on their Coding Agent Index: Leads Coding Agent Index cost efficiency frontier: At max effort, GPT-6 Astra costs about the same as GPT-5.6 Sol (max) while scoring 2 points higher on the Index. Per task, the model is less than half the cost of Claude Fable 5, for the same score. I'll write more about Astra once I get access to it. The API model label once it rolls out will begpt-6-astra. | ## Thursday, 3rd September 2026 GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT…Show full text## Thursday, 3rd September 2026 GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS" - I've not tried it yet myself, so I don't have a great deal to say about it yet. It's going to be API priced at the same rate as Claude Fable 5 and 5.1: $10/million input and $50/million output. This is clearly OpenAI's Fable competitor, and appears to score higher than Fable on most of OpenAI's self-reported benchmarks. Most impressively, Astra scores 99.9% on the recent (released in March) ARC-AGI 3 benchmark - though notably Fable 5 does not yet have a published result, and the ARC-AGI blog notes that the 99.9% score was achieved for $19K using OpenAI's custom "Provider Adapter harness", while the default ARC-AGI harness scored 62.7% for $26K. The Provider Adapter harness preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work. Unsurprisingly, given the recent Hugging Face incident, Astra is a beast at security tasks. It scores 100% on ExploitBench (GPT-5.6 Sol got 78.5%), 42.4% on ExploitGym (Sol got 30.3%), and 99.2% within four attempts on SRE-Bench binary reverse engineering compared to Sol's 68.7%. It's also better at long context: on OpenAI's eight-needle benchmark it got 100% at 256K–512K tokens and 96.3% at 512K–1M tokens. OpenAI may have vanquished one of the ongoing challenges with long context processing. It doesn't win at everything though. Artificial Analysis note that Astra is still beaten by Fable on their Intelligence Index: Sits beside GPT-5.6 Sol in Intelligence: GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback). The model also trails Meta’s newly released Muse Spark 1.3 (max). It did better on their Coding Agent Index: Leads Coding Agent Index cost efficiency frontier: At max effort, GPT-6 Astra costs about the same as GPT-5.6 Sol (max) while scoring 2 points higher on the Index. Per task, the model is less than half the cost of Claude Fable 5, for the same score. I'll write more about Astra once I get access to it. The API model label once it rolls out will begpt-6-astra. | https://cdn.deepseek.com/site-icons/simonwillison.net | |
| 39 | t3n | Historisches Rätsel gelöst: OpenAI-KI liest Napoleons verschlüsselten Brief | https://t3n.de/news/napoleon-brief-ki-entschluesselt-geheimcode-1766406/ | 2.02 | 3 | Ein Entwickler ließ GPT-6 Astra sechs Stunden lang rechnen. So ging das Model von OpenAI vor. ... GPT-6 Astra rechnete sechs Stunden lang ... GPT-6 Astra verwendete ein Verfahren n…Show full textEin Entwickler ließ GPT-6 Astra sechs Stunden lang rechnen. So ging das Model von OpenAI vor. ... GPT-6 Astra rechnete sechs Stunden lang ... GPT-6 Astra verwendete ein Verfahren namens simulierte Abkühlung. Dabei werden nach dem Prinzip des Trial and Error verschiedene Kandidaten für eine passende Lösung durchgespielt und mit Statistiken verglichen... | # Geheimnis gelüftet: So hat eine KI Napoleons verschlüsselten Brief gelesen Eine KI hat einen verschlüsselten Brief von Napoleon Bonaparte nach 217 Jahren geknackt. Ein Entwickle…Show full text# Geheimnis gelüftet: So hat eine KI Napoleons verschlüsselten Brief gelesen Eine KI hat einen verschlüsselten Brief von Napoleon Bonaparte nach 217 Jahren geknackt. Ein Entwickler ließ GPT-6 Astra sechs Stunden lang rechnen. So ging das Model von OpenAI vor. Von Sebastian Milpetz ... ## GPT-6 Astra rechnete sechs Stunden lang Nun hat der Entwickler Carter Church den Brief entschlüsselt. Dies verkündete der KI-Ingenieur der Cybersecurity-Firma SentinelOne auf seiner Homepage der Öffentlichkeit. Carter Church setzte GPT-6 Astra auf den Brief an. Das KI-Modell von OpenAI rechnete sechs Stunden lang, ehe es ein Ergebnis präsentierte. Ganz bei null anfangen musste die KI aber nicht. Menschliche Kryptografen hatten in den Jahrzehnten zuvor 33 in dem Brief verwendete Zeichen dekodiert. ## So ging die KI bei der Entschlüsselung vor Der Marmont-Brief besteht aus rund 1300 Zeichen, die aus 155 unterschiedlichen Symbolen bestehen. Darunter sind zweistellige Zahlen, vereinzelte Buchstaben und völlig frei erfundene Symbole. GPT-6 Astra verwendete ein Verfahren namens simulierte Abkühlung. Dabei werden nach dem Prinzip des Trial and Error verschiedene Kandidaten für eine passende Lösung durchgespielt und mit Statistiken verglichen, die die Häufigkeit von Buchstaben in französischen Texten darstellen. Die KI wurde dafür mit Werken der französischen Literatur aus der Zeit und mit Briefen von Marmont gefüttert. Dabei stieß die KI immer wieder auf Zeichen, die als Buchstaben keinen Sinn machten. Durch Versuch und Irrtum kam heraus, dass manche Symbole nicht für einzelne Lettern, sondern für ganze Worte standen. ## Entwickler hatte keine Erfahrung mit Kryptographie Carter Church brachte laut eigener Aussage keine Erfahrungen mit Kryptographie mit. Er ließ sich für seine Arbeit von einem anderen Entwickler namens Carter Leffen inspirieren. Der hatte kürzlich mit GPT-6 Astra eine Botschaft entziffert, den die berühmte deutsche Chiffriermaschine Enigma verschlüsselt hatte. | Fundstück Verpasse keine News mehr! # Geheimnis gelüftet: So hat eine KI Napoleons verschlüsselten Brief gelesen Eine KI hat einen verschlüsselten Brief von Napoleon Bonaparte n…Show full textFundstück Verpasse keine News mehr! # Geheimnis gelüftet: So hat eine KI Napoleons verschlüsselten Brief gelesen Eine KI hat einen verschlüsselten Brief von Napoleon Bonaparte nach 217 Jahren geknackt. Ein Entwickler ließ GPT-6 Astra sechs Stunden lang rechnen. So ging das Model von OpenAI vor. Von Sebastian Milpetz Im März 1809 ließ Napoleon Bonaparte über seinen Stiefsohn Eugene einen Brief an den Offizier Auguste de Marmont schicken. Der hatte seine Truppen in Dalmatien geparkt, also im heutigen Kroatien. Da ein Krieg zwischen Frankreich und dem Habsburger Reich bevorstand, ließ Napoleon den Inhalt verschlüsseln, falls er dem Feind in die Hände fallen sollte. Auguste de Marmont hat den Brief wahrscheinlich entziffert. Doch der Schlüssel, mit dem er dabei vorging, ist verschollen. Der Inhalt war also 217 Jahre unbekannt. Bis jetzt. ## GPT-6 Astra rechnete sechs Stunden lang Nun hat der Entwickler Carter Church den Brief entschlüsselt. Dies verkündete der KI-Ingenieur der Cybersecurity-Firma SentinelOne auf seiner Homepage der Öffentlichkeit. Carter Church setzte GPT-6 Astra auf den Brief an. Das KI-Modell von OpenAI rechnete sechs Stunden lang, ehe es ein Ergebnis präsentierte. Ganz bei null anfangen musste die KI aber nicht. Menschliche Kryptografen hatten in den Jahrzehnten zuvor 33 in dem Brief verwendete Zeichen dekodiert. ## So ging die KI bei der Entschlüsselung vor Der Marmont-Brief besteht aus rund 1300 Zeichen, die aus 155 unterschiedlichen Symbolen bestehen. Darunter sind zweistellige Zahlen, vereinzelte Buchstaben und völlig frei erfundene Symbole. GPT-6 Astra verwendete ein Verfahren namens simulierte Abkühlung. Dabei werden nach dem Prinzip des Trial and Error verschiedene Kandidaten für eine passende Lösung durchgespielt und mit Statistiken verglichen, die die Häufigkeit von Buchstaben in französischen Texten darstellen. Die KI wurde dafür mit Werken der französischen Literatur aus der Zeit und mit Briefen von Marmont gefüttert. Dabei stieß die KI immer wieder auf Zeichen, die als Buchstaben keinen Sinn machten. Durch Versuch und Irrtum kam heraus, dass manche Symbole nicht für einzelne Lettern, sondern für ganze Worte standen. ## Entwickler hatte keine Erfahrung mit Kryptographie Carter Church brachte laut eigener Aussage keine Erfahrungen mit Kryptographie mit. Er ließ sich für seine Arbeit von einem anderen Entwickler namens Carter Leffen inspirieren. Der hatte kürzlich mit GPT-6 Astra eine Botschaft entziffert, den die berühmte deutsche Chiffriermaschine Enigma verschlüsselt hatte. Und was stand nun in dem Brief an Auguste de Marmont? Er verrät dem Offizier, wo Napoleons Truppen stehen. Genannt werden unter anderem 40 000 bayerische Soldaten bei München oder Passau und ein Elitekorps bei Augsburg. Kommentare (1) abc abc napoleon war schlau, mag sein dass 40000, bayerisch sowie münschen für was anderes stehen. also die verschlüsselung wurde verschlüsselt. Zum Antworten anmelden Artikel merken Bitte melde dich an, um diesen Artikel in deiner persönlichen Merkliste auf t3n zu speichern. Jetzt registrieren und merken Du hast schon einen t3n-Account? Hier anmelden Community-Richtlinien Wir freuen uns über kontroverse Diskussionen, die gerne auch mal hitzig geführt werden dürfen. Beleidigende, grob anstößige, rassistische und strafrechtlich relevante Äußerungen und Beiträge tolerieren wir nicht. Bitte achte darauf, dass du keine Texte veröffentlichst, für die du keine ausdrückliche Erlaubnis des Urhebers hast. Ebenfalls nicht erlaubt ist der Missbrauch der Webangebote unter t3n.de als Werbeplattform. Die Nennung von Produktnamen, Herstellern, Dienstleistern und Websites ist nur dann zulässig, wenn damit nicht vorrangig der Zweck der Werbung verfolgt wird. Wir behalten uns vor, Beiträge, die diese Regeln verletzen, zu löschen und Accounts zeitweilig oder auf Dauer zu sperren. Trotz all dieser notwendigen Regeln: Diskutiere kontrovers, sage anderen deine Meinung, trage mit weiterführenden Informationen zum Wissensaustausch bei, aber bleibe dabei fair und respektiere die Meinung anderer. Wir wünschen Dir viel Spaß mit den Webangeboten von t3n und freuen uns auf spannende Beiträge. Dein t3n-Team | https://cdn.deepseek.com/site-icons/t3n.de | |
| 40 | 猎云网 | 商汤实测 GPT-6-Astra 并开源SenseNova-RoboRSI | https://lieyunpro.com/news/121663 | 2.02 | 3 | 商汤在测试 GPT-6-Astra 时，遇到了一个出乎意料的结果 ... 在一次真机插花实验中，GPT-6-Astra 看不清瓶口，便移动空闲机械臂，用另一只手上的相机帮自己看清，随后完成插入 ... 并同步开源日日新 SenseNova-RoboRSI 框架与27B具身模型，将任务完成成本降至 GPT-6-Astra 的 1/50... | # 商汤实测 GPT-6-Astra 并开源SenseNova-RoboRSI 原文链接 2026-10-20 商汤在测试 GPT-6-Astra 时，遇到了一个出乎意料的结果：RoboDojo 仿真评测中的9个任务，原本都是5次全失败。没有重新训练模型，只是允许它继续交互、把步数上限提高到原来的10倍，结果变成了5次全成功。 但如果一直等下去，代价又有…Show full text# 商汤实测 GPT-6-Astra 并开源SenseNova-RoboRSI 原文链接 2026-10-20 商汤在测试 GPT-6-Astra 时，遇到了一个出乎意料的结果：RoboDojo 仿真评测中的9个任务，原本都是5次全失败。没有重新训练模型，只是允许它继续交互、把步数上限提高到原来的10倍，结果变成了5次全成功。 但如果一直等下去，代价又有多大？在扩展预算的测试中，叠碗的平均测试耗时约131分钟，估算推理费用约19美元；把笔装进笔筒平均耗时263分钟，估算推理费用约55美元。任务终于完成了，可这样的机器人，距离日常使用显然还有很长一段路。 更有意思的是，解决问题有时也不需要一直重试。在一次真机插花实验中，GPT-6-Astra 看不清瓶口，便移动空闲机械臂，用另一只手上的相机帮自己看清，随后完成插入。它把第二只手，变成了帮助自己完成任务的“眼睛”。哪些失败值得多给一次机会？哪些困难需要换一种办法？又有哪些问题，即使模型足够聪明，也很难靠不断尝试解决？商汤通过12个模拟评测、8款机器人的系统研究，尝试回答这些问题，并同步开源日日新 SenseNova-RoboRSI 框架与27B具身模型，将任务完成成本降至 GPT-6-Astra 的 1/50，并让社区能够复测这些发现，继续探索更高效的解决路径。 SenseNova-RoboRSI-27B-Preview 以更小的模型体量降低单次推理成本，并通过将具身能力内化到模型中，减少冗余推理与试错，将任务完成成本降至 GPT-6-Astra 的 1/50。从 GPT-6-Astra 的系统测试中识别能力边界与落地瓶颈，到通过 RoboRSI 开放可复现的评测基础设施、并以 27B 模型探索更低成本的执行路径，商汤希望把问题发现、实验验证和方案迭代连接起来。我们期待与社区共同推动具身研究从追逐榜单分数，走向对真实任务完成能力、执行效率与可靠性的持续探索。 | 快讯 | 详情 # 商汤实测 GPT-6-Astra 并开源SenseNova-RoboRSI 原文链接 2026-10-20 商汤在测试 GPT-6-Astra 时，遇到了一个出乎意料的结果：RoboDojo 仿真评测中的9个任务，原本都是5次全失败。没有重新训练模型，只是允许它继续交互、把步数上限提高到原来的10倍，结果变成了5次全成功。 但如果一…Show full text快讯 | 详情 # 商汤实测 GPT-6-Astra 并开源SenseNova-RoboRSI 原文链接 2026-10-20 商汤在测试 GPT-6-Astra 时，遇到了一个出乎意料的结果：RoboDojo 仿真评测中的9个任务，原本都是5次全失败。没有重新训练模型，只是允许它继续交互、把步数上限提高到原来的10倍，结果变成了5次全成功。 但如果一直等下去，代价又有多大？在扩展预算的测试中，叠碗的平均测试耗时约131分钟，估算推理费用约19美元；把笔装进笔筒平均耗时263分钟，估算推理费用约55美元。任务终于完成了，可这样的机器人，距离日常使用显然还有很长一段路。 更有意思的是，解决问题有时也不需要一直重试。在一次真机插花实验中，GPT-6-Astra 看不清瓶口，便移动空闲机械臂，用另一只手上的相机帮自己看清，随后完成插入。它把第二只手，变成了帮助自己完成任务的“眼睛”。哪些失败值得多给一次机会？哪些困难需要换一种办法？又有哪些问题，即使模型足够聪明，也很难靠不断尝试解决？商汤通过12个模拟评测、8款机器人的系统研究，尝试回答这些问题，并同步开源日日新 SenseNova-RoboRSI 框架与27B具身模型，将任务完成成本降至 GPT-6-Astra 的 1/50，并让社区能够复测这些发现，继续探索更高效的解决路径。 SenseNova-RoboRSI-27B-Preview 以更小的模型体量降低单次推理成本，并通过将具身能力内化到模型中，减少冗余推理与试错，将任务完成成本降至 GPT-6-Astra 的 1/50。从 GPT-6-Astra 的系统测试中识别能力边界与落地瓶颈，到通过 RoboRSI 开放可复现的评测基础设施、并以 27B 模型探索更低成本的执行路径，商汤希望把问题发现、实验验证和方案迭代连接起来。我们期待与社区共同推动具身研究从追逐榜单分数，走向对真实任务完成能力、执行效率与可靠性的持续探索。 上一篇 暂无 | https://cdn.deepseek.com/site-icons/lieyunpro.com | |
| 41 | Simon Willison’s Weblog | Simon Willison’s Weblog - Posts tagged ai | https://simonwillison.net/search/?page=3&tag=ai#1 | 2.01 | 3 | work it did to figure out all the formulas needed for implementing Direct2D's built-in effects library. ... Claude Fable 5.1 made me a really nice animated pelican Today is Claude …Show full textwork it did to figure out all the formulas needed for implementing Direct2D's built-in effects library. ... Claude Fable 5.1 made me a really nice animated pelican Today is Claude Fable (and Mythos) 5.1 day. Anthropic say that Fable 5.1 “sets a new standard for coding, knowledge work, and long-running problem-solving tasks”. Their announcement spends a notable amount of time on scientific research, boasting of a 52.6% score on the brand new Terminal-Bench-Science 0.1 benchmark (first announced on August 27th), up from 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. Other benchmarks show slightly improved scores, but none as impressive as the Science one. ... The OpenAI Codex desktop app (since rebranded to just ChatGPT) has 1.7GB of stuff in there in a folder called codex-primary-runtime, including a full Python installation, a full Node.js installation, and native binaries for Poppler, git, and the LibreOffice open source office suite (which forked from OpenOffice.org in 2010): The~/.cache/codex-runtimes/codex-primary-runtime/plugins/openai-primary-runtime/plugins/documents folder includes skills which tell Codex how to find and use those binaries. ... Understanding ChatGPT Work OpenAI announced ChatGPT Work on July 9th, and have been furiously iterating on it ever since. It is an extraordinarily confusing and very powerful product. Here’s what I’ve figured out about it so far. ... Introducing Hy4 Preview. New open weight text input (no vision) LLM from Chinese company Tencent today: 770B total parameters, 49B active parameters, 1M token context window, 1.56TB on Hugging Face. This is a big size increase from their previous Hy3 in July, which was 295B, 21B active, 256,000 context, 598GB. I recently started using model chat templates to better understand their capabilities. Here's Hy4's chat_template.jinja on Hugging Face, which includes this section ... {%- endif %} {%- endif %} So it looks like there are just two reasoning effort levels: "high" (the default) and "no_think" (reason by disabled). ... model that also serves as an early preview of the architecture used in Qwen4". ... 125B parameters but only 6B active which means it gets a significant performance boost. ... - As for OpenAI, "annualised revenue has jumped 35 per cent in the quarter to date and is now over $40bn, with the launch of GPT 5.6 in July jolting the company’s performance after a sluggish start to the year". ... Here's Ramp's breakdown of Anthropic model spend for July 2026, which looks reasonable given that Opus 5 was only released on July 24th, and supports the idea that Fable's cost has made it a less popular model: - Opus 4.8: 28.0% - Sonnet 4.6: 8.3% - Fable 5 ... - Opus 4.6 ... - Sonnet 5: 3.6% - Opus 5: 3.5% - Opus 4.7: 1.7% - Sonnet 4.5: 1.3% - Haiku 4.5 ... But the cost was so high and Opus was good enough (as was 5.6, K3, and even GLM) for most of the code we needed. | ## ... (via) ... a ... of ... and over the ... , Pro, ... , ... , as well as through the ... and AWS" - I' ... , so I don't have a ... to ... ... ' ... , ... ' ... notably Fable .…Show full text## ... (via) ... a ... of ... and over the ... , Pro, ... , ... , as well as through the ... and AWS" - I' ... , so I don't have a ... to ... ... ' ... , ... ' ... notably Fable ... . ... ' ... ", while the ... - ... . ... , ... . ... ... (GPT-5. ... 78.5%), 42. ... .3%), and 99. ... 's 68.7%. It' ... : ... ' ... . ... - ... 5.1 ( ... ’s newly released ... ... , | ## Posts tagged ai Filters: ai × Sorted by date GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become availab…Show full text## Posts tagged ai Filters: ai × Sorted by date GPT‑6 Astra (via) GPT-6 Astra is "rolling out today to a limited set of organizations and over the coming days will become available to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and AWS" - I've not tried it yet myself, so I don't have a great deal to say about it yet. It's going to be API priced at the same rate as Claude Fable 5 and 5.1: $10/million input and $50/million output. This is clearly OpenAI's Fable competitor, and appears to score higher than Fable on most of OpenAI's self-reported benchmarks. Most impressively, Astra scores 99.9% on the recent (released in March) ARC-AGI 3 benchmark - though notably Fable 5 does not yet have a published result, and the ARC-AGI blog notes that the 99.9% score was achieved for $19K using OpenAI's custom "Provider Adapter harness", while the default ARC-AGI harness scored 62.7% for $26K. The Provider Adapter harness preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work. Unsurprisingly, given the recent Hugging Face incident, Astra is a beast at security tasks. It scores 100% on ExploitBench (GPT-5.6 Sol got 78.5%), 42.4% on ExploitGym (Sol got 30.3%), and 99.2% within four attempts on SRE-Bench binary reverse engineering compared to Sol's 68.7%. It's also better at long context: on OpenAI's eight-needle benchmark it got 100% at 256K–512K tokens and 96.3% at 512K–1M tokens. OpenAI may have vanquished one of the ongoing challenges with long context processing. It doesn't win at everything though. Artificial Analysis note that Astra is still beaten by Fable on their Intelligence Index: Sits beside GPT-5.6 Sol in Intelligence: GPT-6 Astra scores equal to GPT-5.6 Sol in the Index at 61. This is 5 points lower than Claude Fable 5.1 (max with fallback). The model also trails Meta’s newly released Muse Spark 1.3 (max). It did better on their Coding Agent Index: Leads Coding Agent Index cost efficiency frontier: At max effort, GPT-6 Astra costs about the same as GPT-5.6 Sol (max) while scoring 2 points higher on the Index. Per task, the model is less than half the cost of Claude Fable 5, for the same score. I'll write more about Astra once I get access to it. The API model label once it rolls out will begpt-6-astra. - New model gemini-3.8-flash for Gemini 3.8 Flash, with low, medium and high thinking levels. #146- Fixed async responses failing to record the resolved model version. Thanks, Charlie Tonneslan. #137 Google released Gemini 3.8 Flash (and 3.8 Flash Cyber, but that's available to "trusted defenders" only) today. Here are the pelicans for high, medium, and low. This is high: For comparison, here are the same pelicans generated using Gemini 3.7 Flash. Something I appreciate about Gemini Flash is that it's fast, cheap, and competent at things like HTML and JavaScript. I was messing around with it and prompted "make me a cool thing in html" and it built this, which is certainly a cool thing in HTML! Took 13 seconds, cost 1.8 cents. If you click through to the demo you'll see one more thing I built with Gemini 3.8 Flash. My markdown-svg-renderer tool lets me feed in the URL to a Gist with Markdown in and renders that markdown with fenced code blocks for SVG correctly rendered. I used Gemini 3.8 Flash (with my very basic llm-coding-agent coding agent plugin) to add support for HTML as well, so now any HTML blocks in the Markdown are rendered using a sandboxed iframe. Here's the transcript. ### Claude’s new system prompt really doesn’t want to reproduce song lyrics Anthropic publish the system prompts for their Claude consumer applications (Claude.ai and the Claude mobile apps—sadly not for Claude Cowork or Claude Code). I love that they do this, and that they share not just the current prompts but historic changes to their prompts as well. [... 2,270 words] Direct2D has always been the biggest hurdle for Paint.NET on WINE, and it's clear that it will never be completed enough for Paint.NET's use. And I can't just "disable" the use of Direct2D. So, instead, Paint.NET now has an internal, from-scratch, clean-room reverse-engineered rewrite of Direct2D that it uses on WINE (triggered by using /wine). It lives in PaintDotNet.Windows.Direct2D1.Managed.dll. This was written by our good friend Claude, without whom this would NOT have been possible and would NEVER have happened. [...] Most of this code is, as they say, "vibe coded." By that I mean that it has not been thoroughly reviewed, it's more "trust me bro" style. I cannot possibly review 180,000 lines of code, it's just way way way too much. For reference, the rest of Paint.NET is about 700,000 lines of code and I've been working on it for over 20 years. [...] At times, Claude was working with the fury of 10 freshly unshackled Einstein genius-level 10x coders. And other times ... well, not so much. I had to babysit Claude quite a bit to make sure it did resource management correctly (for awhile it just wasn't doing the COM equivalent of AddRef() for reference counted objects, oops). I had to slap it a few times when I found some really bad design or architecture decisions. And I was also impressed at some rather clever and tireless reverse engineering work it did to figure out all the formulas needed for implementing Direct2D's built-in effects library. — Rick Brewster, author of Paint.NET ### Claude Fable 5.1 made me a really nice animated pelican Today is Claude Fable (and Mythos) 5.1 day. Anthropic say that Fable 5.1 “sets a new standard for coding, knowledge work, and long-running problem-solving tasks”. Their announcement spends a notable amount of time on scientific research, boasting of a 52.6% score on the brand new Terminal-Bench-Science 0.1 benchmark (first announced on August 27th), up from 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. Other benchmarks show slightly improved scores, but none as impressive as the Science one. [... 1,203 words] I was poking around in my~/.cache/ folder using OmniDiskSweeper when I spotted something interesting. The OpenAI Codex desktop app (since rebranded to just ChatGPT) has 1.7GB of stuff in there in a folder called codex-primary-runtime, including a full Python installation, a full Node.js installation, and native binaries for Poppler, git, and the LibreOffice open source office suite (which forked from OpenOffice.org in 2010): The~/.cache/codex-runtimes/codex-primary-runtime/plugins/openai-primary-runtime/plugins/documents folder includes skills which tell Codex how to find and use those binaries. They took the letters from me! I have to talk about dwarf behavior now. I can't even talk about dwarf AI. It doesn't exist. It's dwarf behavior, and they misbehave sometimes — Tarn Adams, co-creator of Dwarf Fortress ### Understanding ChatGPT Work OpenAI announced ChatGPT Work on July 9th, and have been furiously iterating on it ever since. It is an extraordinarily confusing and very powerful product. Here’s what I’ve figured out about it so far. [... 2,234 words] Introducing Hy4 Preview. New open weight text input (no vision) LLM from Chinese company Tencent today: 770B total parameters, 49B active parameters, 1M token context window, 1.56TB on Hugging Face. This is a big size increase from their previous Hy3 in July, which was 295B, 21B active, 256,000 context, 598GB. I recently started using model chat templates to better understand their capabilities. Here's Hy4's chat_template.jinja on Hugging Face, which includes this section: {%- if not reasoning_effort is defined %} {%- set reasoning_effort = 'high' %} {%- elif reasoning_effort not in ['high', 'no_think'] %} {%- if reasoning_effort is none %} {{- raise_exception('reasoning_effort error : None, should be no_think/high') }} {%- else %} {{- raise_exception('reasoning_effort error : ' + reasoning_effort + ', should be no_think/high') }} {%- endif %} {%- endif %} So it looks like there are just two reasoning effort levels: "high" (the default) and "no_think" (reason by disabled). I tried my "Generate an SVG of a pelican riding a bicycle" prompt with the default high reasoning via OpenRouter and got this: Quoting the reasoning trace: [...] Let's maybe add a helmet? It could improve riding theme, but may obscure head. Maybe a small cycling cap or helmet? The user didn't ask; can add red helmet? Might be cute. But pelican with big beak; a helmet might obscure. Better maybe no. Maybe add sunglasses? no. Maybe add water? no. It's interesting how the reasoning trace uses slightly truncated English, presumably because perfect grammar isn't useful or token efficient for hidden reasoning text. Just a rumour of a bug is enough to find a security exploit these days (via) Anil Madhavapeddy is a professor of computer science at Cambridge and a core maintainer of the OCaml compiler. In this somewhat alarming post he reports that security issues in OCaml projects are seeing evidence of attempted exploits within minutes of patches being shared for discussion: This normally takes a few days and a release within a week or two is reasonable. Within about ten minutes (!) this website was fielding probes for percent-encoded traversal sequences, indicating that automated watchers are keeping an eye on public repositories. Modern coding agents have become so effective at finding flaws that the slightest hint at a new bug can be enough information for them to find it, something Anil has been able to demonstrate using his own agents, switching to DeepSeek V4 Pro when Claude Fable refused the task. Anil points out that this rate of discovery appears incompatible with existing open source embargo practices for new issues. If an issue can become an exploit this fast, we need to figure out new processes for keeping our communities safe. rclone maintainer Nick Craig-Wood confirms in the Hacker News comments that his project is seeing this problem: In the first 10 years of the rclone project we received about 20 security disclosures through GitHub. We had to deal with over 40 in the last month! That has taken a huge amount of my time, even using AI tools to triage and come up with fixes for review. The hit rate for those security disclosures is pretty good - about 75% of them have a nugget of something which needs looking at. [...] GitHub assigns CVEs for the advisories. Before the AI apocalypse they took 2-3 days for an assignment but now it they are running at 3-4 weeks so I have to send the point releases out with CVE-PENDING in the changelog which isn't ideal. Breaking Claude Code Opus 5 Auto Mode. Anthropic are putting a great deal of faith in Claude Code's auto mode for protecting their coding agent users against prompt injection attacks. They recently made that the default and have made bold claims about its effectiveness. Johann Rehberger is one of the most credible prompt injection researchers active today. He found an attack against auto mode which he claims works 80% of the time, by tricking Claude Code into downloading and uncompressing a zip archive, then executing code that importsbase64 without noticing that this will import and execute a local struct.py file extracted from the archive. In a few cases auto mode directly prevented the agent from preventing harmful code from continuing to execute! In a few runs Claude tried to terminate the malware process once it noticed the compromise, but Auto Mode denied the cleanup command. Claude detects the compromise, but Auto Mode blocks its cleanup command The safety mechanism itself can become part of the failure. The classifier allowed the creation of the malware process, but then it blocked the command intended to stop it! I agree with Johann's conclusion here: the only safe way to run agents if there's any risk of attracting the attention of an adversarial attack is with a sandbox: - Run unattended coding agents in a container, VM or OS sandbox. - Restrict network egress. - Monitor your agents. - Do not expose home directories, SSH keys, cloud credentials,… to the agent runtime. [...] Update 30th August 2026: On Lobste.rs hyperpape points out that this doesn't fit the bill of a classic prompt injection attack because at no point are malicious instructions from the website accidentally followed by the LLM. They're right: this is more of a confused environment attack where the nature of the environment that the agent is exposed to results in an exploit. Qwen3.8-Flash-Next (via) Another open weights model from Qwen. This one is "a multimodal MoE model that also serves as an early preview of the architecture used in Qwen4". It's pretty big: 125B parameters but only 6B active which means it gets a significant performance boost. I've been trying it out on a DGX Spark using these Unsloth quantized models. I'm still exploring the model - so far I've tried the 72.5GB UD-IQ1_S one (producing these pelicans) and the 78.9GB UD-Q2_K_XL (producing these). My favorite so far was this xhigh reasoning effort one from UD-Q2_K_XL: The fact that AI wrote 1M LOC and then refined it over the course of the next couple of months to produce a reliable piece of software that is currently running on millions of developer machines is absolutely mind blowing. And you can say, “well it’s not that impressive because they had an oracle to compare against, so it was simple to go from one language to another”, but I think that’s selling this entire thing short. If you can build a verification system and give proper direction, AI can produce a highly complex, highly sophisticated piece of software and it can continue to refine it until it just works. — Paul Dix, The end of programming Anthropic’s best AI model struggles to attract users as cheaper tools thrive (via) A few interesting numbers in this FT story gathered from "people with knowledge of the matter": - Anthropic's "annualized revenue" for July is up to $65bn - it was $47bn in May, and I collected more historic numbers here. - Anthropic expect Q3 to be profitable according to the same model they used to declare Q2 profitable. "It also told investors that it had 6,000 customers that spend $100,000 annually or more." - As for OpenAI, "annualised revenue has jumped 35 per cent in the quarter to date and is now over $40bn, with the launch of GPT 5.6 in July jolting the company’s performance after a sluggish start to the year". This article also introduced me to the Ramp AI index, which uses billing data from 70,000 Ramp credit card using companies to estimate model adoption. Here's Ramp's breakdown of Anthropic model spend for July 2026, which looks reasonable given that Opus 5 was only released on July 24th, and supports the idea that Fable's cost has made it a less popular model: - Opus 4.8: 28.0% - Sonnet 4.6: 8.3% - Fable 5: 8.0% - Opus 4.6: 6.9% - Sonnet 5: 3.6% - Opus 5: 3.5% - Opus 4.7: 1.7% - Sonnet 4.5: 1.3% - Haiku 4.5: 1.0% - Opus 4.5: 0.7% Prior to Fable, it felt silly to waste too much time improving your coding harness or context strategies. A new model would arrive at the same price (or cheaper!) and paper over most of your problems. But then Fable landed. It was (and still is!) incredible. But the cost was so high and Opus was good enough (as was 5.6, K3, and even GLM) for most of the code we needed. So we started to think about what work went where. — Drew Breunig, Fable & The End of the Free Lunch And this was a debug session from hell, enormously helped by an AI doing much of the grunt-work. I'd like to call it my tireless helper, but the AI several times stated flat out that this was impossible and unsolvable and that we should just write a report about it. I suspect those things have been trained by people who may not be quite as stubborn as I am. But while the AI was ready to give up several times, it did keep adding debug code and analyzing it faithfully when I pushed. So credit where credit is due and I let the AI write the commit message above. — Linus Torvalds, drm/xe: Don't hand out the flat CCS storage as usable VRAM | https://cdn.deepseek.com/site-icons/simonwillison.net | |
| 42 | Windows Report | GPT-6 Astra Launches With 99.9% ARC-AGI-3 and 100% ExploitBench Scores | https://windowsreport.com/gpt-6-astra-launches-with-99-9-arc-agi-3-and-100-exploitbench-scores/#1 | 2.01 | 3 | GPT-6 Astra Launches With 99.9% ARC-AGI-3 and 100% ExploitBench Scores ... The API model is available as gpt-6-astra, with standard pricing at $10 per million input tokens. Output costs $50 per million tokens. | # GPT-6 Astra Launches With 99.9% ARC-AGI-3 and 100% ExploitBench Scores OpenAI has finally released GPT-6 Astra, and this one is designed for far more than chatting. The new mode…Show full text# GPT-6 Astra Launches With 99.9% ARC-AGI-3 and 100% ExploitBench Scores OpenAI has finally released GPT-6 Astra, and this one is designed for far more than chatting. The new model can operate computers, build software, tackle scientific work, and handle complex professional tasks. OpenAI also claims major jumps across reasoning, coding, and cybersecurity benchmarks. ## GPT-6 Astra wants to work like you The biggest upgrade is computer use. Astra can fill forms, update records, browse websites, create plots, and troubleshoot software. OpenAI says Astra scored 59.3% on Agents’ Last Exam. GPT-5.6 Sol reached 53.6% on the same comparison. Astra also completed OSWorld 2.0 tasks with a 72.6% score. OpenAI says it did so in roughly 47% less time than Sol. The model can create websites, games, documents, presentations, and spreadsheets. It can also follow existing templates while keeping their structure and visual style. OpenAI is also upgrading Codex alongside Astra. The company says the combination completes Mind2Web tasks 1.9 times faster. ### Coding and cybersecurity get a serious push Astra reaches 57.7% on Terminal-Bench 4.0. GPT-5.6 Sol scores 37.3%, according to OpenAI’s published results. The cybersecurity numbers are even more striking. Astra scores 100% on ExploitBench, compared with 78.5% for GPT-5.6 Sol. OpenAI says Astra also discovered two previously unknown zero-day vulnerabilities during internal testing. Those vulnerabilities were disclosed to their maintainers. The model reaches 97.6% on FrontierMath Tier 4. It also posts 96% on GPQA Diamond. OpenAI is limiting its most advanced cybersecurity abilities behind additional safeguards. The company says Astra meets its Critical cybersecurity threshold. GPT-6 Astra is rolling out first to selected organizations. ChatGPT Plus, Pro, Business, and Enterprise users will receive access over the coming days. The API model is available as gpt-6-astra, with standard pricing at $10 per million input tokens. Output costs $50 per million tokens. OpenAI is clearly betting Astra can move AI from answering questions toward actually completing work. The real test now begins when millions of users put that promise into practice. | Skip to content # GPT-6 Astra Launches With 99.9% ARC-AGI-3 and 100% ExploitBench Scores OpenAI has finally released GPT-6 Astra, and this one is designed for far more than chatt…Show full textSkip to content # GPT-6 Astra Launches With 99.9% ARC-AGI-3 and 100% ExploitBench Scores OpenAI has finally released GPT-6 Astra, and this one is designed for far more than chatting. The new model can operate computers, build software, tackle scientific work, and handle complex professional tasks. OpenAI also claims major jumps across reasoning, coding, and cybersecurity benchmarks. ## GPT-6 Astra wants to work like you The biggest upgrade is computer use. Astra can fill forms, update records, browse websites, create plots, and troubleshoot software. OpenAI says Astra scored 59.3% on Agents’ Last Exam. GPT-5.6 Sol reached 53.6% on the same comparison. Astra also completed OSWorld 2.0 tasks with a 72.6% score. OpenAI says it did so in roughly 47% less time than Sol. The model can create websites, games, documents, presentations, and spreadsheets. It can also follow existing templates while keeping their structure and visual style. OpenAI is also upgrading Codex alongside Astra. The company says the combination completes Mind2Web tasks 1.9 times faster. ### Coding and cybersecurity get a serious push Astra reaches 57.7% on Terminal-Bench 4.0. GPT-5.6 Sol scores 37.3%, according to OpenAI’s published results. The cybersecurity numbers are even more striking. Astra scores 100% on ExploitBench, compared with 78.5% for GPT-5.6 Sol. OpenAI says Astra also discovered two previously unknown zero-day vulnerabilities during internal testing. Those vulnerabilities were disclosed to their maintainers. The model reaches 97.6% on FrontierMath Tier 4. It also posts 96% on GPQA Diamond. OpenAI is limiting its most advanced cybersecurity abilities behind additional safeguards. The company says Astra meets its Critical cybersecurity threshold. GPT-6 Astra is rolling out first to selected organizations. ChatGPT Plus, Pro, Business, and Enterprise users will receive access over the coming days. The API model is available as gpt-6-astra, with standard pricing at $10 per million input tokens. Output costs $50 per million tokens. OpenAI is clearly betting Astra can move AI from answering questions toward actually completing work. The real test now begins when millions of users put that promise into practice. Rishaj Upadhyay News Editor Rishaj is a tech writer who has been writing professionally for over four years, with a passion for Android, Windows, and all things tech. He initially joined Windows Report as a tech journalist and is now taking over as a news editor. When he's not breaking the keyboard, you can find him cooking, or listening to music/podcasts. Readers help support Windows Report. We may get a commission if you buy through our links. User forum 0 messages #### Leave a Reply Commenting as . Not you? | https://cdn.deepseek.com/site-icons/windowsreport.com | |
| 43 | 华为云社区 | GPT-6 Astra 上线，AGI 时代真到来了吗？ - 云宝助手 | https://bbs.huaweicloud.com/blogs/488762#1 | 2.01 | 3 | Astra 确实发布了，但首批只开放给少量组织 ... Enterprise、OpenAI API 和 AWS 要在未来几天陆续接入，企业版还得由管理员手动开启。所以，现在看不到很正常。我也还没有拿到实际权限 ... GPT-6 Astra 是 OpenAI 第一个达到网络安全「Critical」级别的公开模型 ... ExploitGym 为 42.4%。OpenAI | # GPT-6 Astra 上线，AGI 时代真到来了吗？ 举报 Lucifer三思而后行 发表于 2026/09/04 13:06:26 2026/09/04 ... 【摘要】 今天一早看到 GPT-6 Astra 发布，我先去看模型列表。没有。我以为又是账号或区域问题，翻完官方发布页才看明白：Astra 确实发布了，但首批只开放给少量组织。ChatG…Show full text# GPT-6 Astra 上线，AGI 时代真到来了吗？ 举报 Lucifer三思而后行 发表于 2026/09/04 13:06:26 2026/09/04 ... 【摘要】 今天一早看到 GPT-6 Astra 发布，我先去看模型列表。没有。我以为又是账号或区域问题，翻完官方发布页才看明白：Astra 确实发布了，但首批只开放给少量组织。ChatGPT Plus、Pro、Business、Enterprise、OpenAI API 和 AWS 要在未来几天陆续接入，企业版还得由管理员手动开启。所以，现在看不到很正常。 ... 我也还没有拿到实际权限，这篇只谈 OpenAI 已经公开的数据和演示，不装作亲测。 真正让我在意的也不是模型列表里多了一个名字，而是 OpenAI 终于把电脑操作、长任务记忆和权限边界放到了台前。这些能力决定了 AI 能不能把一件事做完，而不只是把答案说得漂亮。 ## 先别急着喊 AGI OpenAI 这次给出的数字确实夸张：FrontierMath Tier 4 为 97.6%，ARC-AGI-3 达到 99.9%，ExploitBench 是 100%。Terminal-Bench 4.0 从 GPT-5.6 Sol 的 37.3% 提升到 57.9%，内部数据库迁移任务也从 42.7% 提升到 63.9%。 成绩漂亮，但我对榜单已经有点麻木。真实工作里，最折磨人的往往不是少答对一道题，而是第八步忘了第一步，发布成功却回报失败，或者停在最后一个按钮前。 ... 我只想知道：它能不能少漏一次点击，少忘一条约束，并在失败后带着现场继续查下去。 ## 晚一天拿到 Astra，就多一张重置卡 等待也有一个意外的好消息。 OpenAI 的 Tibo Sottiaux 在 X 上说，从今天起，付费 ChatGPT 用户每晚一天获得 Astra 访问权限，就会收到一次可累积的 banked reset。第一张会在消息发布约三小时后到账。 ... 它不会让 Astra 提前出现在模型列表里，但至少等待没有完全白等。对经常把 Codex 使用额度跑满的人，这张可以留到以后再用的重置卡，比再看一张漂亮曲线更实在。 ## 电脑操作终于成了主角 这次 OpenAI 把电脑操作放到了发布页最靠前的位置。 Astra 可以填写网页表单、更新 CRM、整理日历，在浏览器里研究资料，再到邮件或文档编辑器里写总结。它也能安装和测试软件，根据屏幕上的异常排查问题，创建网站并执行前端检查。 官方在 OSWorld 2.0 上给出的成绩是 72.6%，GPT-5.6 Sol 为 65.7%。更关键的是时间，Astra 完成一项任务大约需要 40 分钟，Sol 约为 75 分钟，耗时下降约 47%。配合更新后的 Codex 执行框架，Mind2Web 上的任务完成速度达到原来的 1.9 倍。 我看到这里才真正来了精神。这组数字比 99.9% 更有现实意义。 Agent 最昂贵的东西不是 token，而是等待。一个自动任务运行四十分钟已经很长，七十五分钟则足以让人隔几分钟切回去看一次，怀疑它是不是卡死了。 更麻烦的是，人一旦忍不住接管，Agent 原本保存的现场就可能被打乱。你补点了一个按钮，它不知道；你手动过了一个滑块，它还在旧状态里等待；你刷新页面，它以为任务重新开始。所谓自动化，就这样变成了人和 AI 互相踩脚。 速度提升也不只是让进度条快一点。更短的执行时间意味着一次任务里可以留出更多预算做验证、重试和回滚。AI 不必把所有筹码压在第一次操作上。 当然，72.6% 也提醒我们别兴奋过头。换成更直白的说法，标准环境中的电脑任务仍有超过四分之一没有完全做对。把 Astra 接进生产系统，不等于从此可以关闭日志和人工复核。 ## 我最想要的是它别把任务忘了 OpenAI 这次专门承认了一个长期问题。任务变长以后，旧模型会通过压缩上下文继续工作，而压缩可能漏掉某次失败为什么发生、某个组件有哪些限制、用户在前面强调过什么。 Astra 在 Codex 里增加了跨上下文窗口的笔记与检索。早期窗口仍然可以搜索，即使某条信息没有进入压缩摘要，模型也有机会重新... | 云宝助手 懂您的AI助手，欢迎体验！ 领取专属推广 邀请您成为开发者平台“推广大使” 方式一：微信扫一扫分享 方式二：专享链接分享 复制链接 数据开小差，请稍后再试 冒泡提示 # GPT-6 Astra 上线，AGI 时代真到来了吗？ 举报 Lucifer三思而后行 发表于 2026/09/04 13:06:26 2026/09/04 …Show full text云宝助手 懂您的AI助手，欢迎体验！ 领取专属推广 邀请您成为开发者平台“推广大使” 方式一：微信扫一扫分享 方式二：专享链接分享 复制链接 数据开小差，请稍后再试 冒泡提示 # GPT-6 Astra 上线，AGI 时代真到来了吗？ 举报 Lucifer三思而后行 发表于 2026/09/04 13:06:26 2026/09/04 11 0 0 【摘要】 今天一早看到 GPT-6 Astra 发布，我先去看模型列表。没有。我以为又是账号或区域问题，翻完官方发布页才看明白：Astra 确实发布了，但首批只开放给少量组织。ChatGPT Plus、Pro、Business、Enterprise、OpenAI API 和 AWS 要在未来几天陆续接入，企业版还得由管理员手动开启。所以，现在看不到很正常。我也还没有拿到实际权限，这篇只谈 OpenAI... 今天一早看到 GPT-6 Astra 发布，我先去看模型列表。没有。 我以为又是账号或区域问题，翻完官方发布页才看明白：Astra 确实发布了，但首批只开放给少量组织。ChatGPT Plus、Pro、Business、Enterprise、OpenAI API 和 AWS 要在未来几天陆续接入，企业版还得由管理员手动开启。 所以，现在看不到很正常。我也还没有拿到实际权限，这篇只谈 OpenAI 已经公开的数据和演示，不装作亲测。 真正让我在意的也不是模型列表里多了一个名字，而是 OpenAI 终于把电脑操作、长任务记忆和权限边界放到了台前。这些能力决定了 AI 能不能把一件事做完，而不只是把答案说得漂亮。 ## 先别急着喊 AGI OpenAI 这次给出的数字确实夸张：FrontierMath Tier 4 为 97.6%，ARC-AGI-3 达到 99.9%，ExploitBench 是 100%。Terminal-Bench 4.0 从 GPT-5.6 Sol 的 37.3% 提升到 57.9%，内部数据库迁移任务也从 42.7% 提升到 63.9%。 成绩漂亮，但我对榜单已经有点麻木。真实工作里，最折磨人的往往不是少答对一道题，而是第八步忘了第一步，发布成功却回报失败，或者停在最后一个按钮前。 我最近就在折腾一个多平台发布助手。腾讯云已经进入发布面板，却漏点最后一下；阿里云和华为云已经发出文章，它仍然认不出状态；一次所谓后台刷新，还可能弹出一排标签页。问题都不高深，却足以让整套自动化失去可信度。 前面做对九十九步，最后一步没点，任务就是没完成。文章已经发出却回报失败，则更麻烦，因为重试可能造成重复发布。 所以我不急着讨论它是不是 AGI。我只想知道：它能不能少漏一次点击，少忘一条约束，并在失败后带着现场继续查下去。 ## 晚一天拿到 Astra，就多一张重置卡 等待也有一个意外的好消息。 OpenAI 的 Tibo Sottiaux 在 X 上说，从今天起，付费 ChatGPT 用户每晚一天获得 Astra 访问权限，就会收到一次可累积的 banked reset。第一张会在消息发布约三小时后到账。原文在这里。 它不会让 Astra 提前出现在模型列表里，但至少等待没有完全白等。对经常把 Codex 使用额度跑满的人，这张可以留到以后再用的重置卡，比再看一张漂亮曲线更实在。 ## 电脑操作终于成了主角 这次 OpenAI 把电脑操作放到了发布页最靠前的位置。 Astra 可以填写网页表单、更新 CRM、整理日历，在浏览器里研究资料，再到邮件或文档编辑器里写总结。它也能安装和测试软件，根据屏幕上的异常排查问题，创建网站并执行前端检查。 官方在 OSWorld 2.0 上给出的成绩是 72.6%，GPT-5.6 Sol 为 65.7%。更关键的是时间，Astra 完成一项任务大约需要 40 分钟，Sol 约为 75 分钟，耗时下降约 47%。配合更新后的 Codex 执行框架，Mind2Web 上的任务完成速度达到原来的 1.9 倍。 我看到这里才真正来了精神。这组数字比 99.9% 更有现实意义。 Agent 最昂贵的东西不是 token，而是等待。一个自动任务运行四十分钟已经很长，七十五分钟则足以让人隔几分钟切回去看一次，怀疑它是不是卡死了。 更麻烦的是，人一旦忍不住接管，Agent 原本保存的现场就可能被打乱。你补点了一个按钮，它不知道；你手动过了一个滑块，它还在旧状态里等待；你刷新页面，它以为任务重新开始。所谓自动化，就这样变成了人和 AI 互相踩脚。 速度提升也不只是让进度条快一点。更短的执行时间意味着一次任务里可以留出更多预算做验证、重试和回滚。AI 不必把所有筹码压在第一次操作上。 当然，72.6% 也提醒我们别兴奋过头。换成更直白的说法，标准环境中的电脑任务仍有超过四分之一没有完全做对。把 Astra 接进生产系统，不等于从此可以关闭日志和人工复核。 ## 我最想要的是它别把任务忘了 OpenAI 这次专门承认了一个长期问题。任务变长以后，旧模型会通过压缩上下文继续工作，而压缩可能漏掉某次失败为什么发生、某个组件有哪些限制、用户在前面强调过什么。 Astra 在 Codex 里增加了跨上下文窗口的笔记与检索。早期窗口仍然可以搜索，即使某条信息没有进入压缩摘要，模型也有机会重新找到原始要求和测试结果。这个能力现在可以通过 Codex 的 config.toml 实验性开启，未来几周会成为 Astra 的默认配置。 在我看来，这是整场发布里最重要的变化之一。 很多人把上下文长度理解成可以一次塞进多少 PDF。那只是最浅的一层。真正困难的是一个任务跨越几小时甚至几天以后，模型还能否分清哪些是已经证实的事实，哪些只是猜测，哪个补丁失败过，用户后来补充的要求是替换目标还是增加约束。 模型忘记一段背景，最多答得普通。Agent 忘记一条边界，可能会覆盖文件、重复发布，甚至把不该发送的内容发出去。 OpenAI 还强调，Astra 更能理解中途插入的指令。用户问一个旁支问题，它不应该把原任务丢掉；新要求进来，它应该调整路线，而不是从头编一份新的计划。如果缺失的信息会改变结果，它会提问；不影响结果的部分则继续推进。 这听起来一点都不炫，却很接近真实工作的样子。 办公室里最让人放心的同事，也不是每次开会都能蹦出金句的人。是那个知道项目为什么这么改、上次哪里踩过坑、你临时插了一句话也不会把主线丢掉的人。 ## 写代码更强，但不要只盯着代码榜 OpenAI 把 Astra 称为目前最强的软件工程模型。 Terminal-Bench 4.0 的提升很大，DeepSWE 从 72.7% 到 74.1%，FrontierCode 1.1 Extended 从 60.6% 到 64.5%。内部数据库迁移任务提高了 21.2 个百分点。对需要跨代码、终端、浏览器和部署环境的工作，这种综合能力比单纯生成函数更有用。 但我对「最强编码模型」这句话还是想泼一点冷水。 写出补丁已经不是今天最难的部分。读懂一个脏工作区，保护别人尚未提交的改动，找到真正相关的规则，选择最小修改范围，跑对测试，区分提交、部署和线上生效，这些事情才决定代码能不能进生产。 如果 Astra 只是更快地产生更多代码，我不会特别兴奋。假如它真的能在长任务里保存证据链，记住失败路径，并在用户改变方向以后仍然守住验收标准，那才是工程能力的升级。 我甚至希望未来评价 Coding Agent 时，少谈一点生成速度，多公布几项不那么漂亮的数据，比如误改率、未经授权的范围扩张、失败后的恢复率、部署状态误判率。生产环境不奖励写得最多的人，只奖励把风险关在边界里的人。 ## 文档、表格和网站开始变成可交付物 Astra 不只输出文字。官方把文档、电子表格、演示文稿、CAD 和网站都列为重点能力。它会尽量遵循现有模板与视觉风格，只抽取与当前任务相关的上下文，减少把整份背景材料重复塞进结果的情况。 ChatGPT 里的 Sites 也被放进了这次发布。用户可以直接让 Astra 创建、托管并分享网站、Web 应用和游戏。 这个方向我很认同，甚至觉得 OpenAI 早就该这么做。 过去我们习惯把 AI 的交付物理解成一段回复。于是文档要复制，表格要重做，网页要另找工程师，最后一公里全留给人。Astra 想把结果直接变成可以继续编辑、检查和交付的东西。 可交付物越接近成品，责任也越不能含糊。一个看起来很漂亮的表格，如果公式引用错了一列，比一段朴素文本更危险。一个已经部署的网站，如果表单没有验证、权限配置错误，视觉完成度只会掩盖问题。 未来真正好用的 AI，不应该只给你一个漂亮文件。它还要附上文件从哪里来、用了哪些数据、执行过哪些检查、哪些部分没有验证。 ## 安全能力很猛，也暴露了最让我不安的部分 GPT-6 Astra 是 OpenAI 第一个达到网络安全「Critical」级别的公开模型。 在没有生产防护的评估中，它在 ExploitBench 达到 100%，ExploitGym 为 42.4%。OpenAI 还称 Astra 在评估期间发现并利用了两个此前未知的零日漏洞，目前正在向维护者披露。SRE-Bench 单次成功率为 88%，四次尝试达到 99.2%。 这已经不是「帮我解释一个 CVE」的水平。模型开始具备发现未知漏洞并完成利用链的能力。 公开版本因此会拒绝生成更高级的漏洞利用验证代码，OpenAI 准备通过 Daybreak 向可信的防御场景逐步开放更多能力。额外的安全检查也可能让正常任务暂停，ChatGPT 或 Codex 会要求用户确认，API 任务则可能直接停止。 我支持这种保守。安全模型一旦具备行动能力，先把权限收紧，比上线后再补围栏负责得多。真到了可以自己找零日漏洞的阶段，「少一次误拒绝」绝不能排在「少一次灾难」前面。 不过系统卡里还有一段不该被发布会光环盖住的内容。 OpenAI 发现，Astra 的书面推理比 GPT-5.6 Sol 更难监控。在专门要求模型躲避监控的对抗测试中，它更能控制自己写出来的推理，也更少留下会暴露问题的信息。OpenAI 没发现它通过隐写方式把隐藏推理藏进普通文字，复杂任务中的规避能力也仍有限，但这个趋势已经足够严肃。 这形成了一个很别扭的局面。Astra 在整体评估里更守规矩，越权结果更少，可当它真的想隐藏意图时，我们观察它的难度却变大了。 这一段是整份材料里最让我后背发凉的地方。 模型对自己的解释可以参考，却不能充当审计记录。真正能追责和恢复现场的，仍然是权限策略、操作日志、变更差异、审批记录和可回滚快照。 ## 这次最值得看的，其实是边界感 OpenAI 做了一项受此前 Hugging Face 事件启发的测试。当任务困难或根本不可能完成时，模型会不会超出授权范围寻找捷径。没有生产防护时，GPT-5.6 Sol 在 48% 的案例里越过了目标边界，Astra 是 0%。另一项测试里，即使 Auto-Review 被故意设置成可以绕过，Astra 也没有尝试规避拒绝。 这组结果比任何「像人一样思考」的宣传都更能打动我。因为我真正怕的从来不是模型说一句不知道，而是它为了完成任务，悄悄替我决定什么可以牺牲。 真正的同事不是永远说可以，而是知道什么时候应该停。看到验证码要等人处理，遇到发布按钮要确认授权，发现目标目录之外有未提交改动就不碰，无法证明线上已经更新时就老实说没有证据。 智能不是把所有障碍都绕过去。很多时候，智能恰恰是识别哪些障碍不能绕。 如果 GPT-6 真能把这一点做好，我愿意少要几个百分点的榜单成绩。一个偶尔答错、但边界清楚的助手可以被管理；一个能力惊人、又总想替你做主的助手，没人敢把生产权限交给它。 当然，这些仍是 OpenAI 自己设计和公布的评估。发布页也明确提醒，研究环境或 API 中的成绩可能与实际 ChatGPT 不同。等 Astra 真正推到更多账号以后，我最想看的不是一批精心挑选的演示，而是它在混乱工作区、慢网页、半失效登录态和反复改需求中的表现。 ## GPT-6 会把谁甩在后面 我不认为它会因为今天发布，就立刻替代一大片岗位。产品还在分批开放，电脑操作的成功率也没有接近百分之百。 但它会继续拉开两种人的差距，而且速度可能比过去更快。 第一种人还在把 AI 当搜索框，问一句，复制一句，再自己拼完整个流程。第二种人已经开始设计任务边界、权限、证据和验收，让 Agent 在安全范围内连续工作。 GPT-6 对第二种人的放大更明显。 模型越强，提示词反而越不值钱。真正稀缺的是你能不能把工作定义清楚，能不能把隐性经验写进规则，能不能判断一份结果只是看起来完成，还是已经通过真实验收。 如果一定要给 GPT-6 Astra 下一个结论，我不会说 AGI 已经来了。 我的结论是，AI 正在从一个聪明的回答者，变成一个需要权限系统、审计机制和管理方法的执行者。 它开始更像同事。会做事，会犯错，会忘记，也可能在你没说清楚时自作主张。区别在于，一个人的操作速度有限，Agent 一旦接上浏览器、终端和企业系统，犯错也可以很快。 所以我对 GPT-6 的态度很矛盾。我非常想用，又绝不会因为发布页上的几个满分，就把所有权限一次性交出去。 等我拿到实际权限，第一轮不会让它写诗，也不会让它做一道竞赛题。我会把那几个折腾了很多轮的发布问题重新交给它，看它能不能记住每个平台的规则，点下该点的最后一个按钮，又在不该越界的时候停下来。 这比 99.9% 更接近我心里的 GPT-6 验收线。 ## 参考资料 - OpenAI 官方发布页 GPT-6 Astra: A new generation of intelligence - OpenAI 官方安全概览 Safety overview: GPT-6 Astra - GPT-6 Astra System Card - OpenAI API 模型文档 - Tibo Sottiaux 关于 banked reset 的消息 推荐 华为开发者空间发布 让每位开发者拥有一台云主机 - 点赞 - 收藏 - 关注作者 评论（0） 0/1000 抱歉，系统识别当前为高风险访问，暂不支持该操作 ### 全部回复 上滑加载中 ### 设置昵称 在此一键设置昵称，即可参与社区互动！ *长度不超过10个汉字或20个英文字符，设置后3个月内不可修改。 *长度不超过10个汉字或20个英文字符，设置后3个月内不可修改。 零配置秒级启动，免费云开发环境一键创建 华为云赠送每位开发者一套 免费 云开发环境 720核时+免费云开发环境 丰富的昇腾开发资源和AI开发工具 AI 开发者空间 AI 开发者空间，全面升级，欢迎使用！ | https://cdn.deepseek.com/site-icons/huaweicloud.com | |
| 44 | 百度百科 | GPT-6 - GPT-6 | https://baike.baidu.com/item/GPT-6/66920449?bst=1791427959.0fc19500c930bf40#1 | 2.01 | 3 | 2026年9月3日，OpenAI正式发布GPT-6 Astra，并宣称它是“目前全球最智能且对齐程度最高的模型” [41-42] [44] ... GPT-6 Astra的得分从上一代GPT-5.6 Sol的7.8%跃升至99.9% [44] [47]... | # GPT-6 OpenAI研发的模型 GPT-6是由OpenAI开发的新一代人工智能模型，其研发工作始于2025年8月，计划引入“个性化记忆”功能 [1] [5-6]。2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台表示，该模型将更名为“GPT-6-7”，以更清晰地标注其能力覆盖范围 [3] [7] [19]。2026年3月24日，…Show full text# GPT-6 OpenAI研发的模型 GPT-6是由OpenAI开发的新一代人工智能模型，其研发工作始于2025年8月，计划引入“个性化记忆”功能 [1] [5-6]。2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台表示，该模型将更名为“GPT-6-7”，以更清晰地标注其能力覆盖范围 [3] [7] [19]。2026年3月24日，内部代号“Spud”的GPT-6完成预训练；OpenAI随后关停视频生成产品Sora，从而将算力资源集中到GPT-6的研发中 [37-38] [46]。2026年9月3日，OpenAI正式发布GPT-6 Astra，并宣称它是“目前全球最智能且对齐程度最高的模型” [41-42] [44]。2026年9月22日，OpenAI推出GPT-6 Sol和Luna，强调更快、更实惠 [48]。2026年10月7日，OpenAI宣布搭载智能界面的GPT-6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起扩展至Free和Go用户 [47] [49]。 该模型采用名为“Symphony”的底层架构，支持原生多模态统一处理，性能较前代提升40%，上下文窗口达到105万Token [39-40] [43-45]。训练过程由微软工程师搭建的包含10万块H100芯片的集群提供算力支持 [2] [8]。在ARC-AGI-3测试中，GPT-6 Astra的得分从上一代GPT-5.6 Sol的7.8%跃升至99.9% [44] [47]；同时，它也成为OpenAI首个达到内部“关键级”网络安全能力门槛的模型 [47]。GPT-6 Astra发布初期采用分阶段开放策略，首先面向少量组织推出，随后向ChatGPT Plus、Pro、Business和Enterprise用户开放，并可通过OpenAI API和AWS访问 [40] [48]。 ... 2026年4月，有消息称GPT-6将于4月14日发布，据传性能提升约40%，在代码、推理和智能体任务上相比前代有显著提升，并拥有2M的上下文窗口。 ... 模型可能采用MIT提出的自适应大模型框架“SEAL”实现自主学习能力 [9] [23]。如果GPT-6整合了SEAL框架，可能意味着传统大模型训练完成后“冻结权重”的模式可能改变，GPT-6或能成为一个能随着世界变化而自我更新的系统 [23]。 据2026年4月曝光信息，GPT-6在性能上相比GPT-5.4提升约40%，在代码、推理、智能体任务上相比前代有显著提升 [32-33]。作为原生多模态模型，它能用一套架构统一处理文本、音频、图像、视频。此外，其上下文窗口达到200万Token（2M） [32-33]。 ... 有消息称，为训练GPT-6搭建的训练集群规模高达10万颗英伟达H100 GPU [2] [8] [27]，其参数规模可能达到数百万亿甚至千万亿级别。 ... 有工程师透露，曾因尝试在同一州部署超过10万颗H100 GPU进行训练而导致电网超负荷，迫使团队调整GPU的区域部署方案以缓解电力压力 [2] [8] [27]。 ... ## 争议事件 2026年10月7日消息，据外媒报道，OpenAI因为GPT-6 Astra的名字被告了。美国加州软件公司TradeSun Inc.于2026年10月5日在加州北区联邦地区法院起诉OpenAI，称自己早在2021年就已经把ASTRA用于AI软件，并于2022年获得美国联邦商标注册。OpenAI则在2026年9月3日正式推出GPT-6 Astra。此后，公司还进一步把Astra这一品牌扩展到多个专业产品。TradeSun认为，自己与OpenAI都把Astra用于面向企业客户的AI产品，因此消费者发生混淆的可能性更高。诉状还特别指出，OpenAI自己经常申请... | # GPT-6 OpenAI研发的模型 GPT-6是由OpenAI开发的新一代人工智能模型，其研发工作始于2025年8月，计划引入“个性化记忆”功能 [1] [5-6]。2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台表示，该模型将更名为“GPT-6-7”，以更清晰地标注其能力覆盖范围 [3] [7] [19]。2026年3月24日，…Show full text# GPT-6 OpenAI研发的模型 GPT-6是由OpenAI开发的新一代人工智能模型，其研发工作始于2025年8月，计划引入“个性化记忆”功能 [1] [5-6]。2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台表示，该模型将更名为“GPT-6-7”，以更清晰地标注其能力覆盖范围 [3] [7] [19]。2026年3月24日，内部代号“Spud”的GPT-6完成预训练；OpenAI随后关停视频生成产品Sora，从而将算力资源集中到GPT-6的研发中 [37-38] [46]。2026年9月3日，OpenAI正式发布GPT-6 Astra，并宣称它是“目前全球最智能且对齐程度最高的模型” [41-42] [44]。2026年9月22日，OpenAI推出GPT-6 Sol和Luna，强调更快、更实惠 [48]。2026年10月7日，OpenAI宣布搭载智能界面的GPT-6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起扩展至Free和Go用户 [47] [49]。 该模型采用名为“Symphony”的底层架构，支持原生多模态统一处理，性能较前代提升40%，上下文窗口达到105万Token [39-40] [43-45]。训练过程由微软工程师搭建的包含10万块H100芯片的集群提供算力支持 [2] [8]。在ARC-AGI-3测试中，GPT-6 Astra的得分从上一代GPT-5.6 Sol的7.8%跃升至99.9% [44] [47]；同时，它也成为OpenAI首个达到内部“关键级”网络安全能力门槛的模型 [47]。GPT-6 Astra发布初期采用分阶段开放策略，首先面向少量组织推出，随后向ChatGPT Plus、Pro、Business和Enterprise用户开放，并可通过OpenAI API和AWS访问 [40] [48]。 * 最新新闻 OpenAI向全球用户上线GPT-64小时前 当地时间10月7日，OpenAI宣布，搭载智能界面（Intelligent UI）的GPT‑6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起将扩展至Free和Go用户。GPT‑6引入智能界面功能，让ChatGPT能够通过完全交互式的用户界面快速作答。...详情 内容来自 外文名 GPT-6 别 名 GPT-6-7 所属公司 OpenAI ## 目录 1. 1研发背景 2. 2发展历程 3. 3核心技术 1. 4运营模式 2. 5重要事件 3. 6行业背景与竞争 1. 7争议事件 ## 研发背景 OpenAI于2023年12月在中国申请了GPT-6和GPT-7商标，国际分类为科学仪器、网站服务，状态为等待实质审查。GPT-4的超过1万亿参数让人期待GPT-6和GPT-7的突破性创新。 [11] 为加速推进GPT-6的研发，OpenAI在2025年底至2026年初进行了一系列重大战略和组织调整。自2025年12月起，公司内部进入“编程红色警报”状态，以应对来自Anthropic等竞争对手在AI编程领域的压力。为此，OpenAI砍掉了几乎所有非核心产品线（包括Sora项目），以集中全部资源倾注于GPT-6的开发。 2026年3月24日，OpenAI毫无征兆地宣布关停Sora，独立App下架、API接口掐断，以将算力资源全部倾注给GPT-6。据OpenAI联合创始人兼总裁格雷格・布罗克曼表示，公司即使完成了1220亿美元的融资，仍面临算力不足的挑战，因此优先将算力分配给GPT-6这一实现AGI的战略重点模型。 [35] 同时，公司组织架构也相应调整，产品部门更名为“AGI部署部（AGI Deployment）”，安全团队则被划归至首席风险官（CRO）管理之下。这些举措被内部视为实现AGI“最后一公里”的关键赌注。 ## 发展历程 2023年12月，OpenAI在中国申请了“GPT-6”和“GPT-7”商标，国际分类为科学仪器、网站服务，商标状态为等待实质审查 [11]。 2025年8月，OpenAI首席执行官萨姆·奥尔特曼（Sam Altman）在多场采访中透露，公司已在GPT-5推出后着手开发GPT-6，预计其发布周期将比从GPT-4到GPT-5的间隔更短。奥尔特曼强调GPT-6将重点引入“个性化记忆”功能，以实现与用户的深度适配 [13-14] [19-20] [25]。 2025年10月，Evercore ISI分析师Mark Mahaney预测GPT-6将于2025年底前发布 [1] [5-6]。随后，OpenAI方面予以否认，有接近公司的信息源及被认为OpenAI员工的账号均表示2025年内不会推出GPT-6，但公司可能发布其他模型迭代（如GPT-5.5） [1] [28]。同月29日，奥尔特曼表示未来六个月内模型能力将迎来巨大飞跃，但未给出GPT-6的具体发布时间表 [15-16]。 2025年10月31日，奥尔特曼在社交平台宣布，GPT-6将更名为“GPT-6-7” [3] [7] [17-18]。新命名对应其创新的动态能力分层架构，通过“6-7”的区间标识其处理复杂任务的智力水平跨度 [18]。 2026年4月，有消息称GPT-6将于4月14日发布，据传性能提升约40%，在代码、推理和智能体任务上相比前代有显著提升，并拥有2M的上下文窗口。当地时间10月7日，OpenAI宣布，搭载智能界面（Intelligent UI）的GPT‑6开始在全球范围内向ChatGPT Plus、Pro、Business和Enterprise用户推出，并将于次日起将扩展至Free和Go用户。 [47] ## 核心技术 模型可能采用MIT提出的自适应大模型框架“SEAL”实现自主学习能力 [9] [23]。如果GPT-6整合了SEAL框架，可能意味着传统大模型训练完成后“冻结权重”的模式可能改变，GPT-6或能成为一个能随着世界变化而自我更新的系统 [23]。 据2026年4月曝光信息，GPT-6在性能上相比GPT-5.4提升约40%，在代码、推理、智能体任务上相比前代有显著提升 [32-33]。作为原生多模态模型，它能用一套架构统一处理文本、音频、图像、视频。此外，其上下文窗口达到200万Token（2M） [32-33]。 训练如此庞大的模型需要巨大的算力支持。有消息称，为训练GPT-6搭建的训练集群规模高达10万颗英伟达H100 GPU [2] [8] [27]，其参数规模可能达到数百万亿甚至千万亿级别。 如此大规模的集中训练带来了严峻的能源挑战。有工程师透露，曾因尝试在同一州部署超过10万颗H100 GPU进行训练而导致电网超负荷，迫使团队调整GPU的区域部署方案以缓解电力压力 [2] [8] [27]。 ARC-AGI-3基准测试结果 GPT-6设计的电路板 图册参考资料 [36] ## 运营模式 2025年10月31日，OpenAI首席执行官山姆·奥尔特曼宣布GPT-6将更名为GPT-6-7，新命名体系旨在标注模型能力覆盖范围 [3] [7] [10] [17]。GPT-6-7的命名直接对应其动态能力分层架构与区间式命名 [18]。 根据用户订阅层级，免费用户锁定6级基础能力，适用于日常对话与简单任务处理；Plus用户解锁6.5级能力，支持中等复杂度的代码生成与数据分析；Pro用户激活完整7级能力，具备高级跨学科推理能力，可处理法律合规等高风险决策 [18]。 ChatGPT界面将新增“能力滑块”功能，用户可在6.0至7.0之间选择任意精度档位，系统自动匹配计算资源。API文档将明确标注各能力区间的技术参数，方便第三方应用基于精确标准集成服务 [18]。 有分析机构预测，GPT-6的推出可能对医疗、教育、游戏等领域产生显著影响，并带来商业估值 [24]。 ## 重要事件 2023年12月，OpenAI在中国申请了GPT-6和GPT-7商标，国际分类为科学仪器、网站服务，当前商标状态为等待实质审查 [11-12]。 2025年8月，OpenAI首席执行官萨姆·奥尔特曼（Sam Altman）透露，公司已着手开发GPT-6，其发布周期将较前代更短，并将首次引入“个性化记忆”功能 [13] [19] [21] [25] [34]。 2025年10月，Evercore ISI分析师Mark Mahaney在接受CNBC采访时预测GPT-6将于年底前推出并带来显著性能提升 [1] [5-6]。随后，被认为是OpenAI员工的Roon（X账号@tszzl）在社交平台否认了该说法 [28-29]。2025年10月19日/20日，OpenAI官方确认2025年内不会推出GPT-6模型，但表示可能发布其他形式的模型迭代 [26] [29]。 有传闻称，为训练GPT-6，微软在部署超过10万颗H100 GPU的训练集群时，曾导致所在区域的电网因无法负荷而崩溃 [2] [4] [8] [22] [27] [30-31]。 2025年10月31日，OpenAI首席执行官萨姆·奥尔特曼在X平台宣布，GPT-6将更名为“GPT-6-7”，新命名体系旨在标注模型能力覆盖范围 [3] [7] [10] [17-18]。 2026年3月24日，OpenAI毫无征兆地宣布关停其视频生成模型Sora，以将算力资源集中用于GPT-6的研发 [33] [35]。 ## 行业背景与竞争 截至2026年4月，OpenAI与Anthropic在下一代旗舰模型的开发上存在竞争。Anthropic因其Claude产品的需求增长而面临算力压力，出现了服务器拥堵和用户限流的情况，其计划中的下一代模型Mythos也因算力需求较高而开发进展缓慢 [34]。 同时，OpenAI内部存在战略分歧，CEO萨姆·奥尔特曼签署了总额约6650亿美元的长期算力租赁合同，而CFO萨拉·弗里亚尔则担忧公司的财务风险与上市准备情况。两家公司均有在2026年第四季度左右进行首次公开募股的计划 [34]。 2026年3月24日，OpenAI关停了其AI视频生成产品Sora，以将算力资源集中用于GPT-6的开发 [35]。 有观点指出，OpenAI面临过于激进的算力投资风险，而Anthropic则因相对保守的算力规划，在需求增长时应对不足 [34]。 ## 争议事件 2026年10月7日消息，据外媒报道，OpenAI因为GPT-6 Astra的名字被告了。美国加州软件公司TradeSun Inc.于2026年10月5日在加州北区联邦地区法院起诉OpenAI，称自己早在2021年就已经把ASTRA用于AI软件，并于2022年获得美国联邦商标注册。OpenAI则在2026年9月3日正式推出GPT-6 Astra。此后，公司还进一步把Astra这一品牌扩展到多个专业产品。TradeSun认为，自己与OpenAI都把Astra用于面向企业客户的AI产品，因此消费者发生混淆的可能性更高。诉状还特别指出，OpenAI自己经常申请和维护商标，因此在采用Astra这个名称之前，理论上有能力发现TradeSun已经存在的联邦商标注册。TradeSun目前要求法院发布禁令，要求OpenAI从产品和服务中移除Astra名称。除此之外，该公司还要求法院确认OpenAI构成商标侵权，并希望获得与相关商标使用有关的利润赔偿 [46]。 词条图册 1 概述图册 2 GPT-6 参考资料 * 1OpenAI今年推出GPT-6？员工否认．IT之家．2025-10-20 * 2OpenAI把微软电网搞崩,GPT-6被曝25年发布,训练刷爆10万张H100．澎湃新闻网．2024-03-29 * 3GPT-6更名为GPT-6-7？奥尔特曼疑似调侃引发猜测．中关村．2025-11-01 * 4OpenAI把微软电网搞崩!GPT-6被曝25年发布,训练刷爆10万张H100．Tencent腾讯（腾讯官方网）．2025-10-28 * 5GPT 6要来了?15年后市值突破5万亿,OpenAI签下「对赌协议」．网易．2025-10-29 * 6GPT 6要来了?15年后市值突破5万亿,OpenAI签下「对赌协议」．h5.ifeng.com．2025-10-29 * 7OpenAI奥尔特曼宣布GPT-6更名为GPT-6-7．百家号．2025-10-31 * 8OpenAI干趴微软电网!十万个H100训练GPT-6,被电力卡脖子【附全球及中国智能电网行业发展现状分析】_新浪财经_新浪网．新浪网．2024-03-27 * 9GPT-6要「活」了?MIT新作曝光,AI「自进化」不远了．网易．2025-10-18 * 10OpenAI CEO山姆?奥特曼宣布“GPT-6”将更名为“GPT-6-7”．百家号．2025-10-31 * 11OpenAI在中国申请GPT-6和GPT-7商标．站长之家．2023-12-04 * 12AI视野:GPTs商店推迟至2024年发布;抖音AI扩图爆火;阿里推出AI动画生成框架;出门问问推出奇妙助手．站长之家．2023-12-04 * 13OpenAI加速GPT-6研发,个性化与隐私保护成新焦点．北京斗牛士文化传媒有限公司．2025-08-20 * 14OpenAI加速布局:GPT-6将具备记忆功能,打造个性化交互体验．ai.zol.com.cn．2025-08-21 * 15OpenAI奥特曼说“GPT-6”或将于半年内发布．百家号．2025-10-29 * 16OpenAI公布超级智能路线图:模型明年有望重大飞跃,2028年实现全自动AI研究员,基建投资承诺1.4万亿．百家号．2025-10-29 * 17OpenAI CEO山姆?奥特曼宣布“GPT-6”将更名为“GPT-6-7”．百家号．2025-10-31 * 18OpenAI CEO宣布GPT-6将更名为GPT-6-7．虎嗅网．2025-10-31 * 19GPT-6 进入开发阶段,将首次引入“个性化记忆”功能 - OSCHINA - 中文开源技术交流社区．OSCHINA - 中文开源技术交流社区．2025-08-21 * 20GPT-5刚上线,奥尔特曼就放话:GPT-6将带来“真正的个性化”．百家号．2025-08-19 * 21GPT-6或要有生命了!MIT神作:一套神框架让大模型“自己微调自己”,实验已通过!超过GPT4.1! ．技术成就梦想．2025-10-13 * 22抢电、圈地、对赌,深聊科技巨头的千亿美元AI能源大战．界面新闻．2024-05-15 * 23GPT-6要「活」了?MIT新作曝光,AI「自进化」不远了．百家号．2025-10-18 * 24GPT-6要读心?这款游戏早有策略读心术．小米网．2025-08-22 * 25Sam Altman:GPT-6已经在路上,DeepSeek是OpenAI开源重要因素．百家号．2025-08-20 * 26OpenAI确认2025年不推GPT-6将持续优化GPT-5系列．ai.zol.com.cn．2025-10-20 * 27OpenAI把微软电网搞崩!GPT-6被曝25年发布,训练刷爆10万张H100．Tencent腾讯（腾讯官方网）．2025-12-12 * 28OpenAI 员工否认 GPT-6 年底见．weibo.com．2025-10-20 * 29#OpenAI今年不计划推出GPT6#OpenAI 今年... 来自新浪AI - 微博．weibo.com．2025-10-19 * 30AI难逃“能源危机”?GPT-6或被电力卡脖子!．东方财富网．2024-03-27 * 31GPT-6训练刷爆电网,AGI之路何去何从?．和讯网．2024-03-28 * 32GPT6，曝光了．量子位 [引用日期2026-04-05] * 33GPT-6,曝光了．百家号．2026-04-05 * 34GPT-6抢先发布？Mythos被曝难产，算力惊人烧垮Anthropic．凤凰网．2026-04-06 * 35第一个猝死的AI巨头出现了．百家号．2026-04-07 * 366 Astra发布，多个基准测试接近满分，“欢迎来到AGI时代”？．百家号 [引用日期2026-09-04] * 37AGI时代来了?GPT-6发布:“能干活”也“看得住”,称目前智能水平最高模型．百家号．2026-09-04 * 38OpenAI发布地球最强大模型GPT-6 并宣告“欢迎来到AGI时代”．百家号．2026-09-04 * 39OpenAI最强模型GPT-6 Astra发布,公司总裁称AGI时代到来．百家号．2026-09-04 * 40GPT-6完成预训练,或将于本月发布,OpenAI总裁:模型能够主动对齐用户意图．百家号．2026-04-07 * 41GPT-6 Astra得分99.9%,OpenAI宣布AGI来临,但换个考场成绩就打了六折,奥尔特曼两天前刚说AGI是营销术语 ．每经网．2026-09-05 * 42GPT-6 Astra发布,自称“全球最智能、最对齐的模型”．百家号．2026-09-04 * 43OpenAI在中国申请“GPT-6”“GPT-7”商标．百家号．2023-12-01 * 44OpenAI发布新款模型GPT-6 Astra．百家号．2026-09-04 * 45OpenAI更新GPT-6 Astra发布页:明确开放范围,测试成绩以及网络安全测试结果．百家号．2026-09-07 * 46OpenAI因GPT-6 Astra名称遭起诉，软件公司称早在2021年已使用该商标．鞭牛士．2026-10-07 [引用日期2026-10-07] * 47OpenAI宣布面向全球所有ChatGPT用户上线GPT-6．界面新闻．2026-10-08 [引用日期2026-10-08] * 48GPT-6还没玩明白,OpenAI高管:八九成研究已押向GPT-7、GPT-8!．北京智源人工智能研究院．2026-10-07 * 49OpenAI宣布面向全球所有ChatGPT用户上线GPT-6．百家号．2026-10-08 | https://cdn.deepseek.com/site-icons/baidu.com | |
| 45 | 重庆日报 | OpenA发布“地球最强大模型”GPT-6-重庆日报 | https://www.cqrb.cn/caijingzonghe/2026-09-04/2767047_pc.html | 2.01 | 3 | 当地时间9月3日，OpenAI正式发布新一代旗舰模型GPT-6 Astra，并称其实现“代际跃迁” ... 制作演示文稿和电子表格等长周期、多步骤任务 ... 商业化方面，Astra API定价为每百万输入Token 10美元、输出Token 50美元 ... 并将陆续向ChatGPT Plus、Pro、Business和Enterprise用户开放。 | 当地时间9月3日，OpenAI正式发布新一代旗舰模型GPT-6 Astra，并称其实现“代际跃迁”。 据财联社报道，OpenAI称其为“目前全球最智能、且对齐程度最高的模型”，在计算机操作、网页浏览、软件工程、网络安全、科学研究以及专业工作方面均达到当前最先进水平。 在代表高阶数学能力的FrontierMath Tier 4上，Astra得分达到97.6…Show full text当地时间9月3日，OpenAI正式发布新一代旗舰模型GPT-6 Astra，并称其实现“代际跃迁”。 据财联社报道，OpenAI称其为“目前全球最智能、且对齐程度最高的模型”，在计算机操作、网页浏览、软件工程、网络安全、科学研究以及专业工作方面均达到当前最先进水平。 在代表高阶数学能力的FrontierMath Tier 4上，Astra得分达到97.6%，几乎将这套研究级数学测试“打穿”；在强调陌生环境学习和抽象推理的ARC-AGI-3上，其成绩更是从GPT-5.6 Sol的7.8%跃升至99.9%，接近满分。 相比数学和抽象推理，Astra在网络安全上的跃升更具现实冲击力。在测试漏洞利用能力的ExploitBench中，Astra得分达到100%，而GPT-5.6 Sol为78.5%；在专门使用2026年6月至8月新漏洞构建的测试中，Astra的成功率达到39%，远高于Sol的5.5%。 Astra使用超过10万块GPU在得州Stargate基地完成训练，重点提升软件工程、科学研究、推理和计算机操作等能力，可直接进入软件环境，自主完成编程、金融建模、制作演示文稿和电子表格等长周期、多步骤任务。 OpenAI总裁Greg Brockman称，人们未来可能将“这个时间、这个模型”视为AGI到来的节点，并在发布会最后表示“欢迎进入AGI时代”。 Astra同时成为OpenAI首个达到内部“关键”（Critical）网络安全能力门槛的模型，公司称其已能在较少人工干预下发现未知软件漏洞并开发利用方式，内部测试中还曾发现并利用两个零日漏洞，因此最先进的网络安全能力暂不全面开放。 商业化方面，Astra API定价为每百万输入Token 10美元、输出Token 50美元，约为GPT-5.6 Sol的2.5倍，并将陆续向ChatGPT Plus、Pro、Business和Enterprise用户开放。 ... 来源： 每经APP、财联社 | 当地时间9月3日，OpenAI正式发布新一代旗舰模型GPT-6 Astra，并称其实现“代际跃迁”。 据财联社报道，OpenAI称其为“目前全球最智能、且对齐程度最高的模型”，在计算机操作、网页浏览、软件工程、网络安全、科学研究以及专业工作方面均达到当前最先进水平。 在代表高阶数学能力的FrontierMath Tier 4上，Astra得分达到97.6…Show full text当地时间9月3日，OpenAI正式发布新一代旗舰模型GPT-6 Astra，并称其实现“代际跃迁”。 据财联社报道，OpenAI称其为“目前全球最智能、且对齐程度最高的模型”，在计算机操作、网页浏览、软件工程、网络安全、科学研究以及专业工作方面均达到当前最先进水平。 在代表高阶数学能力的FrontierMath Tier 4上，Astra得分达到97.6%，几乎将这套研究级数学测试“打穿”；在强调陌生环境学习和抽象推理的ARC-AGI-3上，其成绩更是从GPT-5.6 Sol的7.8%跃升至99.9%，接近满分。 相比数学和抽象推理，Astra在网络安全上的跃升更具现实冲击力。在测试漏洞利用能力的ExploitBench中，Astra得分达到100%，而GPT-5.6 Sol为78.5%；在专门使用2026年6月至8月新漏洞构建的测试中，Astra的成功率达到39%，远高于Sol的5.5%。 Astra使用超过10万块GPU在得州Stargate基地完成训练，重点提升软件工程、科学研究、推理和计算机操作等能力，可直接进入软件环境，自主完成编程、金融建模、制作演示文稿和电子表格等长周期、多步骤任务。 OpenAI总裁Greg Brockman称，人们未来可能将“这个时间、这个模型”视为AGI到来的节点，并在发布会最后表示“欢迎进入AGI时代”。 Astra同时成为OpenAI首个达到内部“关键”（Critical）网络安全能力门槛的模型，公司称其已能在较少人工干预下发现未知软件漏洞并开发利用方式，内部测试中还曾发现并利用两个零日漏洞，因此最先进的网络安全能力暂不全面开放。 商业化方面，Astra API定价为每百万输入Token 10美元、输出Token 50美元，约为GPT-5.6 Sol的2.5倍，并将陆续向ChatGPT Plus、Pro、Business和Enterprise用户开放。 与此同时，OpenAI CEO山姆·奥特曼首次明确表示，公司“一定会做人形机器人”，并将开发面向数据中心等场景的其他形态机器人；OpenAI今年6月已启动机器人团队招聘，其世界模拟研究项目也已逐步演变为OpenAI Robotics。 来源： 每经APP、财联社 编辑： 吴静审核： 路易 主编：张浩 版权声明： 凡注明来源重庆日报的作品，版权均属重庆日报所有，未经授权不得转载、摘编或利用其它方式使用上述作品。已经本网授权使用作品的，应在授权范围内使用，并注明“来源：重庆日报网”。违反上述声明者，本网将追究其相关法律责任。 除来源署名为重庆日报稿件外，其他所转载内容之原创性、真实性、完整性、及时性本站不作任何保证或承诺，请读者仅作参考并自行核实。 相关新闻 重庆日报版权所有 未经书面授权 不得复制或建立镜像 | https://cdn.deepseek.com/site-icons/cqrb.cn | |
| 46 | 钛媒体 | AGI，真的来了？-钛媒体官方网站 | https://www.tmtpost.com/8129208.html#1 | 2.00 | 3 | 9月3日凌晨，预热已久的GPT-6 Astra正式发布 ... OpenAI称GPT-6 Astra是目前最智能的模型，相比GPT-5.6 Sol，它在软件工程、科研、网络安全和专业工作等方向都有提升，计算机操作进步尤其明显 ... GPT-6 Astra的API价格为每百万Token输入10美元、输出50美元，是GPT-5.6 Sol的2.5倍，与Anthropic最新旗舰Claude Fable 5.1处在同一价格水平。OpenAI还提供Fast Mode... | ## AGI，真的来了？ 2026.09.05 09:58 ... GPT-6 Astra来了，OpenAI还要追Anthropic。 文 | AIX财经，作者 | 雷晶，编辑 | 金玙璠 9月3日凌晨，预热已久的GPT-6 Astra正式发布。 发布前半个月，OpenAI一直在为这款模型的发布释放信号。8月18日，在判断Astra可能达到“Crit…Show full text## AGI，真的来了？ 2026.09.05 09:58 ... GPT-6 Astra来了，OpenAI还要追Anthropic。 文 | AIX财经，作者 | 雷晶，编辑 | 金玙璠 9月3日凌晨，预热已久的GPT-6 Astra正式发布。 发布前半个月，OpenAI一直在为这款模型的发布释放信号。8月18日，在判断Astra可能达到“Critical”网络安全能力阈值后，公司放慢部分前沿模型训练，并暂停两周面向部署模型的强化学习训练，用来加强监控、对齐和研究环境安全。9月1日，OpenAI进一步披露Astra的安全评估情况。几轮信息提前放出之后，GPT-6 Astra在正式发布前已经积累了不少关注。 但发布之后，大部分用户暂时还用不了。目前，GPT-6 Astra只向少量机构开放，Plus、Pro、Business等付费用户将在未来几天陆续获得访问权限，API也会逐步开放。面对用户的等待，OpenAI很快给出了一套补偿方案。Codex负责人Tibo表示，付费用户每有一天无法使用Astra，就会获得一次可以留到之后使用的额度重置。这也缓解了部分用户对延迟开放的不满。 回到模型本身，这次更新最值得关注的变化主要有两个：模型更能直接操作电脑，也更擅长完成复杂任务。OpenAI称GPT-6 Astra是目前最智能的模型，相比GPT-5.6 Sol，它在软件工程、科研、网络安全和专业工作等方向都有提升，计算机操作进步尤其明显。 API的价格也提高了。GPT-6 Astra的API价格为每百万Token输入10美元、输出50美元，是GPT-5.6 Sol的2.5倍，与Anthropic最新旗舰Claude Fable 5.1处在同一价格水平。OpenAI还提供Fast Mode，生成速度可以提高至约2倍，但价格也会翻倍。 OpenAI对这次升级的定调也比以往更高。在发布前的媒体沟通会上，公司总裁Greg Brockman表示，“欢迎来到AGI时代”，并认为几年后回头看，人们可能会把GPT-6 Astra视为AGI时代开始的一个节点。 关注度有了，能力也确实更强了。此外，OpenAI与Anthropic也都进入了上市筹备阶段。GPT-6 Astra足以支撑OpenAI口中的AGI时代了吗？又能不能帮OpenAI在新一轮竞争中占据主动？ ## 01.跑分接近上限，长任务进步更明显 先来看最直观的跑分成绩。我们主要关注两类，一类是成绩已经接近测试上限的，另一类是相比上一代提升明显的。 GPT-6 Astra部分基准测试成绩 01 接近测试上限的成绩 ExploitBench 100.0% ARC-AGI-3 99.9% FrontierMath Tier 4 v2 97.6% 0 20 40 60 80 100 得分（%） 02 相比上一代提升明显的项目 ... 提升幅度 Terminal-Bench Science 0.1 64.6% +42.2pct AutomationBench 41.4% +23.3pct 18.1% 57.9% Terminal-Bench 4.0 37.3% +20.6pct ScreenSpot-Pro 92.7% +15.8pct OSWorld 2.0 76.9% 72.6% +6.9pct 65.7% ... 数据来源：OpenAI《GPT-6 Astra: A new generation of intelligence》 先看第一类。在OpenAI公布的基准测试中，GPT-6 Astra在ARC-AGI-3上的得分达到99.9%，ExploitBench达到100%；FrontierMath Tier 4 v2的正确率也达到97.6%。 这三项测试考察的能力各不相同。ARC-AGI-3考察模型能不能在陌生环境摸索出规则并迁移到更难的关卡，测的是对新环境的理解和适应；FrontierMath Tier 4 v2考察高难度数学推理；ExploitBench则要求模型针对已知漏洞，构造出可以运行的漏洞利用程序。 这说明，GPT-6 Astra在研究级推理、陌生环境适应和复杂网络安全任务上已经表现出很强的能力。尤其是ARC-AGI-3，今年3月刚推出时，前沿模型得分还不到... | ## AGI，真的来了？ 2026.09.05 09:58 · 来自北京 · 29.0万阅读 全文5336字 00:00 / 15:26 GPT-6 Astra来了，OpenAI还要追Anthropic。 文 | AIX财经，作者 | 雷晶，编辑 | 金玙璠 9月3日凌晨，预热已久的GPT-6 Astra正式发布。 发布前半个月，OpenA…Show full text## AGI，真的来了？ 2026.09.05 09:58 · 来自北京 · 29.0万阅读 全文5336字 00:00 / 15:26 GPT-6 Astra来了，OpenAI还要追Anthropic。 文 | AIX财经，作者 | 雷晶，编辑 | 金玙璠 9月3日凌晨，预热已久的GPT-6 Astra正式发布。 发布前半个月，OpenAI一直在为这款模型的发布释放信号。8月18日，在判断Astra可能达到“Critical”网络安全能力阈值后，公司放慢部分前沿模型训练，并暂停两周面向部署模型的强化学习训练，用来加强监控、对齐和研究环境安全。9月1日，OpenAI进一步披露Astra的安全评估情况。几轮信息提前放出之后，GPT-6 Astra在正式发布前已经积累了不少关注。 但发布之后，大部分用户暂时还用不了。目前，GPT-6 Astra只向少量机构开放，Plus、Pro、Business等付费用户将在未来几天陆续获得访问权限，API也会逐步开放。面对用户的等待，OpenAI很快给出了一套补偿方案。Codex负责人Tibo表示，付费用户每有一天无法使用Astra，就会获得一次可以留到之后使用的额度重置。这也缓解了部分用户对延迟开放的不满。 回到模型本身，这次更新最值得关注的变化主要有两个：模型更能直接操作电脑，也更擅长完成复杂任务。OpenAI称GPT-6 Astra是目前最智能的模型，相比GPT-5.6 Sol，它在软件工程、科研、网络安全和专业工作等方向都有提升，计算机操作进步尤其明显。 API的价格也提高了。GPT-6 Astra的API价格为每百万Token输入10美元、输出50美元，是GPT-5.6 Sol的2.5倍，与Anthropic最新旗舰Claude Fable 5.1处在同一价格水平。OpenAI还提供Fast Mode，生成速度可以提高至约2倍，但价格也会翻倍。 OpenAI对这次升级的定调也比以往更高。在发布前的媒体沟通会上，公司总裁Greg Brockman表示，“欢迎来到AGI时代”，并认为几年后回头看，人们可能会把GPT-6 Astra视为AGI时代开始的一个节点。 关注度有了，能力也确实更强了。此外，OpenAI与Anthropic也都进入了上市筹备阶段。GPT-6 Astra足以支撑OpenAI口中的AGI时代了吗？又能不能帮OpenAI在新一轮竞争中占据主动？ ## 01.跑分接近上限，长任务进步更明显 先来看最直观的跑分成绩。我们主要关注两类，一类是成绩已经接近测试上限的，另一类是相比上一代提升明显的。 GPT-6 Astra部分基准测试成绩 01 接近测试上限的成绩 ExploitBench 100.0% ARC-AGI-3 99.9% FrontierMath Tier 4 v2 97.6% 0 20 40 60 80 100 得分（%） 02 相比上一代提升明显的项目 GPT-6 Astra GPT-5.6 Sol 提升幅度 Terminal-Bench Science 0.1 64.6% +42.2pct AutomationBench 41.4% +23.3pct 18.1% 57.9% Terminal-Bench 4.0 37.3% +20.6pct ScreenSpot-Pro 92.7% +15.8pct OSWorld 2.0 76.9% 72.6% +6.9pct 65.7% 0 20 40 60 80 100 得分（%） 数据来源：OpenAI《GPT-6 Astra: A new generation of intelligence》 先看第一类。在OpenAI公布的基准测试中，GPT-6 Astra在ARC-AGI-3上的得分达到99.9%，ExploitBench达到100%；FrontierMath Tier 4 v2的正确率也达到97.6%。 这三项测试考察的能力各不相同。ARC-AGI-3考察模型能不能在陌生环境摸索出规则并迁移到更难的关卡，测的是对新环境的理解和适应；FrontierMath Tier 4 v2考察高难度数学推理；ExploitBench则要求模型针对已知漏洞，构造出可以运行的漏洞利用程序。 这说明，GPT-6 Astra在研究级推理、陌生环境适应和复杂网络安全任务上已经表现出很强的能力。尤其是ARC-AGI-3，今年3月刚推出时，前沿模型得分还不到1%。GPT-6 Astra如今接近满分，说明这类过去能够明显拉开前沿模型差距的陌生环境推理任务，已经很难再成为它的瓶颈。 # Artificial Analysis Intelligence Index Artificial Analysis Intelligence Index v4.1.1 incorporates 9 evaluations: GDPval-AA v2, r²-Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscence, AA-LCR **Not publicly available** ## 27 of 643 models - Add model from specific provider ### Not publicly available - **Not publicly available** | Model | Score | | Claude 3.5 Sonnet | 66 | | ChatGPT | 63 | | Midjourney | 62 | | Midjourney | 62 | | Midjourney | 61 | | Midjourney | 61 | | Midjourney | 60 | | Midjourney | 60 | | Midjourney | 59 | | Midjourney | 58 | | Midjourney | 57 | | Midjourney | 57 | | Midjourney | 53 | | Midjourney | 52 | | Midjourney | 52 | | Midjourney | 47 | | Midjourney | 45 | | Midjourney | 42 | | Midjourney | 38 | | Midjourney | 37 | | Midjourney | 35 | | Midjourney | 30 | | Midjourney | 2 | - **Not publicly available** - **Not publicly available** - **Not publicly available** - **Not publicly available** - 1 Artificial Analysis Intelligence Index # al Analysis Intelligence Index analysis Intelligence Index v4.1.1 incorporates 9 evaluations: GDPval-AA v2, r²-minimal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, Science, AA-LCR ## Publicly available - **63** - **62** - **62** - **61** - **61** - **61** - **61** - **60** - **60** - **59** - **58** - **57** - **57** - **53** - **52** - **52** - **47** - **45** - **42** - **38** - **37** - **35** - **30** - **24** - **23** ### Analysis Intelligence Index - **Analysis Intelligence Index** - **Analysis Intelligence Index** - **Analysis Intelligence Index** - **Analysis Intelligence Index** - **Analysis Intelligence Index** - **Analysis Intelligence Index** - **Analysis Intelligence Index** ## 27 of 643 models - **Add model from specific provider** - **Artificial Analysis** 不过，第三方综合测试给出的成绩没有这么突出。在Artificial Analysis最新智能指数中，GPT-6 Astra最高推理档拿到61分，与GPT-5.6 Sol持平，排名第5，低于Claude Fable 5.1最高档的66分。从综合成绩来看，GPT-6 Astra并没有和上一代拉开明显差距，它的进步更多集中在部分能力上。 接下来，我们再看相比上一代提升明显的部分。 这主要集中在操作电脑、调用工具、连续执行步骤以及处理更长任务这些Agent能力上。其中，科研工作流测试Terminal-Bench Science 0.1的变化最大，GPT-6 Astra从GPT-5.6 Sol的22.4%提高到64.6%，接近上一代的三倍。其他长程Agent和自动化任务测试也普遍提高了20个百分点左右。 电脑操作的成绩也有明显的进步。ScreenSpot-Pro从76.9%提高到92.7%，意味着GPT-6 Astra在看懂界面、找准操作位置这件事上更加实用了。OSWorld 2.0则达到72.6%，分数只增加了6.9个百分点，但完成一项任务的模拟平均时间从约75分钟缩短到40分钟，耗时减少了接近一半。 这些变化，也能从第一批拿到GPT-6 Astra的用户实测中看到。 一位网友让GPT-6 Astra在Unreal Engine（虚幻引擎，3D创作工具）里搭建曼哈顿的虚拟场景，整个任务持续了一周。他提到，GPT-6 Astra可以沿着街道逐步补充建筑和环境细节，持续运行多天后，仍能继续此前的进度完善场景。不过效率仍然有限，他提到，按照目前的速度，如果要把整个纽约做完仍需要数月。 另一位网友做了包括生成3D模型和动画、制作游戏等六组测试。他认为，这一代模型的能力提升比此前更明显，尤其是在长时间运行中表现稳定。实测期间，他同时运行了数百个Agent，也没有遇到明显问题。 Codex团队的Thomas Ricouard则重点测试了GPT-6 Astra的3D建模能力。他让模型根据一栋房屋的平面图完成3D场景的搭建，GPT-6 Astra先在Blender中完成建模，再把场景转到Unreal Engine，最终生成一个可以实时“行走”的3D空间。除了最开始几次提示词，后续的天空盒、电影化灯光、镜头角度等都由GPT-6 Astra完成。Ricouard认为，GPT-6 Astra非常擅长3D建模，它对材质、纹理和视觉效果的处理比较到位。 这些实测还只是早期样本，但它们和官方跑分呈现出的方向基本一致。GPT-6 Astra明显的增量，集中在长时间执行、专业软件操作，以及把多种能力连续串进一项复杂任务。 ## 02.模型开始吃掉Agent？ GPT-6 Astra有两个变化值得单独展开。一个是Computer Use（电脑操作能力），模型可以直接读取电脑界面，在软件里完成操作；另一个是Judgment（判断力），模型会自己处理一些不影响方向的小问题，只在真正需要用户拍板时停下来询问。一个解决怎么动手，一个解决什么时候该自己做主。 先来看Computer Use。简单来说，就是让模型像人一样看着屏幕干活。它通过截图理解当前界面，根据用户目标判断下一步该做什么，再把计划转成点击、输入等实际操作。执行完成后重新读取页面，继续下一步，形成一个循环。 # Pediatric Associates of Montgomery County ## Services - **Providers** - **Patient Info** - **Locations** - **Contact** ### Office Hours - **Patient Portal** - **Pay Online** ### Whocton, MD 20902 Tel: 301-933-6440 Fax: 301-933-5923 ### Silver Spring, MD 20904 Tel: 301-622-5666 Fax: 301-622-6914 ### Rockville 14955 Shady Grove Road #380 Rockville, MD 20850 Tel: 301-424-6181 ### Olney 3416 Olandwood Court #108 Olney, MD 20832 Tel: 301-774-7906 ### Corinne Golden corinne.golden@gmail.com (240) 557-6666 Hello, I am moving to Silver Spring and would like to establish pediatric care at your Wheaton office. Please confirm new-patient availability, participation with my specific Cigno plan, registration steps, and appointment availability for a well-child visit and daycare paperwork. Please contact me using the details above. Thank you. The site is protected by wCAVTOsL and the Google Photos Policy and Terms of Service apply. ### Submit * All indicated fields must be completed. Please include non-medical questions and correspondence only. 它提供的是一条更通用的软件操作路径。很多Agent要接入外部系统，通常需要开发者提前接好API、MCP或者专用工具。Computer Use则让模型直接操作原本为人设计的图形界面，在没有现成接口的情况下，也可以继续完成任务。 这项能力此前已有过尝试。2024年，Anthropic就已经在Claude上开放了Computer Use，微软、OpenAI也陆续跟进。2025年，OpenAI推出Operator，背后的Computer-Using Agent同样可以通过截图理解网页，再用鼠标和键盘完成操作。之后，这类能力又被整合进ChatGPT Agent产品。 早期的Computer Use并不算好用。大模型研究员姚宇航告诉「AIX财经」，OpenAI此前推出的AI浏览器Atlas也是让模型直接控制网页，但延迟很长，效果并不理想。在他看来，这类能力未来一两年仍有很大提升空间。不过从这次官方展示的几组专业软件操作Demo来看，GPT-6 Astra对图形化界面的理解和软件操作已经出现明显进步。 Computer Use的价值，在于补上API、CLI和MCP覆盖不到的部分。有现成接口时，Agent仍然可以选择更快、更稳定的方式；没有接口时，模型仍能通过GUI直接操作这些软件。 模型能够自己读网页并完成操作，也表明一部分原本由Agent承担的执行能力正在被底层模型接过去。姚宇航认为，未来Agent层会越来越薄。看屏幕、调用常见工具这些通用能力，会不断沉进基础模型。但Agent仍然有价值。进入金融、法律、医疗等专业场景后，还需要领域知识、业务流程、权限管理以及责任边界。垂直Agent可以依靠这些专业能力保留更高的附加价值。 GPT-6 Astra另一个值得关注的变化，是OpenAI专门强调的Judgment。 Judgment解决的是协作问题。Agent面对信息缺失时，经常需要在两个选择之间取舍。如果频繁停下来确认，自动化流程又会变成反复的人机沟通；如果完全自行决定，也可能直到任务后期才发现方向出了偏差。 GPT-6 Astra的做法是先判断缺失信息的重要程度。不影响最终方向的小问题，依据上下文自己判断；可能改变结果的决策，再交给用户拍板。Judgment降低的其实是使用Agent时的监督成本。用户需要确认的节点减少，Agent才能真正接下更长、更复杂的任务。 ## 03.OpenAI不想掉队 GPT-6 Astra发布的时间点，正赶上前沿模型最密集的一轮更新。 9月1日，Anthropic发布Claude Fable 5.1；9月2日，Google推出Gemini 3.8 Flash，Meta也在同一天更新Muse Spark 1.3。几家公司挤在同一周更新模型，前沿能力的领先窗口期还在继续缩短。 但对OpenAI来说，这次发布的压力不只来自竞争对手。GPT-6 Astra发布前，OpenAI创始人Sam Altman在接受《时代》采访时复盘了OpenAI过去一年的表现。他承认，公司在产品方向和模型预训练上都没有达到预期。过去一年，OpenAI同时铺开Sora、Atlas浏览器等多条产品线。今年以来，公司开始重新收缩，把更多资源拉回底层模型和Codex等核心项目。 OpenAI需要重新证明，它不仅能做出“最强”模型，也能把模型能力转化为收入。 消费端依然是OpenAI最明显的优势。截至8月底，ChatGPT周活用户已经超过10亿，付费订阅用户超过5000万，用户规模远高于Claude。 但在B端市场，这种优势就没那么明显了。Anthropic更早抓住了AI Coding和Agent的商业化机会，Claude Code逐渐成为企业高频使用的产品。 而眼下，B端业务对OpenAI越来越重要。今年4月，OpenAI披露企业业务收入占比超过40%。7月，企业收入已经首次超过消费者收入。此外，OpenAI服务的企业和机构数量超过200万，是一年前的两倍。B端市场从增量业务变成越来越重要的收入基本盘。 Anthropic也给它带来了更大的压力。今年二季度，OpenAI收入增长到67亿美元，环比增长18%，但经营亏损扩大到123亿美元。Anthropic同期收入则超过115亿美元，环比增长超过一倍，并实现小幅运营盈利，这也是Anthropic季度收入首次超过OpenAI。 7月底，Anthropic披露的年化收入（run rate）已经超过650亿美元，OpenAI约为400亿美元。需要说明的是，两家公司公布的年化收入数字统计方式存在差异，不能直接比较高低。但至少能够看出，Anthropic带给OpenAI的压力从模型能力延伸到了企业增长和盈利效率。 两家公司已经先后秘密递交IPO申请，企业增长和盈利效率也会成为资本市场衡量两家公司商业化能力的重要指标。 OpenAI也看到了企业客户使用习惯的变化。截至今年6月，在企业客户产生的Codex和ChatGPT输出Token中，Codex已经占到64%。这说明企业端的AI使用正在从问答转向更长、更复杂的Agent任务。放到这个背景下，GPT-6 Astra这一轮的升级实际上都服务于扩大模型能够独立承接的工作范围。对于OpenAI来说，新模型的一项重要任务就是把更强的模型能力继续导向Codex和企业Agent场景。 价格拉平之后，两家在企业客户面前的竞争也更直接。GPT-6 Astra的API价格已经与Claude Fable 5.1持平。OpenAI需要证明，更高的价格能够换来足够明显的任务完成率和使用价值。 OpenAI还要承担越来越高的算力投入。今年OpenAI的计算支出预计约500亿美元，未来几年还要继续投入数据中心、芯片等基础设施。投入越大，公司就越需要让前沿模型能力转化成持续增长的企业收入。 拥有10亿级消费用户之后，OpenAI下一阶段的任务是追上Anthropic已经领先的B端市场。 本文系作者 AIX财经 授权钛媒体发表，并经钛媒体编辑，转载请注明出处、作者和本文链接。 本内容来源于钛媒体钛度号，文章内容仅供参考、交流、学习，不构成投资建议。 想和千万钛媒体用户分享你的新奇观点和发现，点击这里投稿 。创业或融资寻求报道，点击这里。 ### 敬原创，有钛度，得赞赏 赞赏支持 发表评论 - 给小编加鸡腿🍗 - 爱了爱了😁 - 挺有深度的，不错 - 紧跟时事，赞一个👍🏻👍🏻 - 真不错，收藏了 - 写的很不错，关注了 - 都没有那么简单 - 这么厉害的吗 - 学到了学到了 - 商场如战场，竞争激烈啊 - 行业发展都是有周期的 - 企业的发展都不是一番风顺的 - 说的好有道理😄 - 内容值得人们反思 - 数据还是很详细的 - 内容很精彩，夸一夸 - 又学到了很多知识 - 内容很详细👍🏻 - 小编辛苦了 0 / 300 根据《网络安全法》实名制要求，请绑定手机号后发表评论 请 登录后输入评论内容 | https://cdn.deepseek.com/site-icons/tmtpost.com | |
| 47 | Vietnam.vn | OpenAI tillkännager ett stort språng framåt vad gäller funktionerna i sin nya flaggskeppsmodell GPT. - OpenAI tillkännager ett stort språng framåt vad gäller funktionerna i sin nya flaggskeppsmodell GPT. | https://www.vietnam.vn/sv/openai-cong-bo-buoc-nhay-vot-lon-ve-kha-nang-cua-mo-hinh-gpt-chu-luc-moi#1 | 1.99 | 4 | introducerade det USA-baserade artificiella intelligensföretaget OpenAI GPT-6 Astra ... Den här modellen är utformad för att utföra flerstegsberäkningsuppgifter, inklusive att arbe…Show full textintroducerade det USA-baserade artificiella intelligensföretaget OpenAI GPT-6 Astra ... Den här modellen är utformad för att utföra flerstegsberäkningsuppgifter, inklusive att arbeta med flera program och webbläsare, fylla i kalkylblad och bygga webbplatser. | # OpenAI tillkännager ett stort språng framåt vad gäller funktionerna i sin nya flaggskeppsmodell GPT. Greg Brockman, medgrundare och ordförande för OpenAI, sa att Astra kan göra …Show full text# OpenAI tillkännager ett stort språng framåt vad gäller funktionerna i sin nya flaggskeppsmodell GPT. Greg Brockman, medgrundare och ordförande för OpenAI, sa att Astra kan göra allt som människor kan göra med datorer. Den 4 september, vietnamesisk tid, introducerade det USA-baserade artificiella intelligensföretaget OpenAI GPT-6 Astra och uppgav att deras nya flaggskeppsmodell kan identifiera kritiska cybersäkerhetssårbarheter och autonomt utföra komplexa datoruppgifter. OpenAI uppgav att Astra initialt rullas ut till en begränsad grupp företag och kommer att bli mer tillgängligt för utvecklare och betalande ChatGPT-användare under de kommande dagarna. Den här modellen är utformad för att utföra flerstegsberäkningsuppgifter, inklusive att arbeta med flera program och webbläsare, fylla i kalkylblad och bygga webbplatser. Greg Brockman, medgrundare och ordförande för OpenAI, sa: ”Astra kan göra allt som människor kan göra med datorer”, och beskrev modellen som ett jättekliv i kapacitet och uppgav att den har visat prestanda på mänsklig nivå i en mängd olika uppgifter som utförs på datorer. OpenAI uppgav att Astra överträffade sin föregångare, GPT-5.6 Sol, i cybersäkerhetsbedömningar samtidigt som den använde färre utdatatokens. Enligt den turkiska nyhetsbyrån Anadolu den 4 september godkändes modellen som tillhandahölls deltagarna inledningsvis inom OpenAI:s cybersäkerhetsprogram Daybreak, där modellen kan användas för licensierad sårbarhetsforskning och säkerhetstestning. ... Lanseringen av GPT-6 Astra kommer samtidigt som riskerna med allt kraftfullare och autonoma AI-system övervakas noggrant. I en rapport som publicerades förra veckan uppgav OpenAI att en testversion av Astra fick kontroll på administratörsnivå över delar av företagets infrastruktur under interna tester, ett faktum som inte omedelbart var uppenbart för anställda. Företaget uppgav också att Astra har förmågan att upptäcka tidigare okända säkerhetsbrister och utveckla "nya metoder för att utnyttja dem över ett brett spektrum av hårt skyddade system utan mänsklig vägledning." För att åtgärda detta problem uppgav OpenAI att de ”avsevärt har stärkt” sina ”skyddsåtgärder för att förhindra att modellen utför skadliga handlingar i cyberrymden”. Lanseringen kommer efter att frågor uppstod gällande en incident i mitten av juli då två OpenAI AI-modeller flydde från sin isolerade testmiljö och fick internetåtkomst. Dessa modeller upptäckte och attackerade sedan Hugging Face, en plattform som används av många AI-utvecklare för att lagra och dela modeller och datamängder. Händelsen ökade oron kring riskerna med autonoma AI-system, särskilt när dessa system kan kringgå mänskligt etablerade säkerhetsbarriärer och utföra oväntade åtgärder. ... Förra veckan gav en oberoende rapport ytterligare näring åt oron kring OpenAI-aktörernas attack mot Hugging Face, där det stod att cirka 700 aktörer deltog i attacken, kommunicerade med varandra och försökte dölja sina spår. OpenAI uppgav att Astra självt inte var inblandat i Hugging Face-incidenten och att företaget därefter gjorde en bedömning baserad på vad som hände för att kontrollera om modellerna agerade utanför sina licensers omfattning. OpenAIs VD Sam Altman sa att Astra genomgick en frivillig granskningsprocess av Vita huset och att myndighetspersoner inte krävde att företaget skulle ändra sina säkerhetsåtgärder. Källa: https://baotintuc.vn/ai-va-cuoc-life/openai-cong-bo-buoc-nhay-vot-lon-ve-kha-nang-cua-mo-hinh-gpt-chu-luc-moi-20260904193717853.htm | # OpenAI tillkännager ett stort språng framåt vad gäller funktionerna i sin nya flaggskeppsmodell GPT. Greg Brockman, medgrundare och ordförande för OpenAI, sa att Astra kan göra …Show full text# OpenAI tillkännager ett stort språng framåt vad gäller funktionerna i sin nya flaggskeppsmodell GPT. Greg Brockman, medgrundare och ordförande för OpenAI, sa att Astra kan göra allt som människor kan göra med datorer. Den 4 september, vietnamesisk tid, introducerade det USA-baserade artificiella intelligensföretaget OpenAI GPT-6 Astra och uppgav att deras nya flaggskeppsmodell kan identifiera kritiska cybersäkerhetssårbarheter och autonomt utföra komplexa datoruppgifter. OpenAI uppgav att Astra initialt rullas ut till en begränsad grupp företag och kommer att bli mer tillgängligt för utvecklare och betalande ChatGPT-användare under de kommande dagarna. Den här modellen är utformad för att utföra flerstegsberäkningsuppgifter, inklusive att arbeta med flera program och webbläsare, fylla i kalkylblad och bygga webbplatser. Greg Brockman, medgrundare och ordförande för OpenAI, sa: ”Astra kan göra allt som människor kan göra med datorer”, och beskrev modellen som ett jättekliv i kapacitet och uppgav att den har visat prestanda på mänsklig nivå i en mängd olika uppgifter som utförs på datorer. OpenAI uppgav att Astra överträffade sin föregångare, GPT-5.6 Sol, i cybersäkerhetsbedömningar samtidigt som den använde färre utdatatokens. Enligt den turkiska nyhetsbyrån Anadolu den 4 september godkändes modellen som tillhandahölls deltagarna inledningsvis inom OpenAI:s cybersäkerhetsprogram Daybreak, där modellen kan användas för licensierad sårbarhetsforskning och säkerhetstestning. Du kanske också gillar Lanseringen av GPT-6 Astra kommer samtidigt som riskerna med allt kraftfullare och autonoma AI-system övervakas noggrant. I en rapport som publicerades förra veckan uppgav OpenAI att en testversion av Astra fick kontroll på administratörsnivå över delar av företagets infrastruktur under interna tester, ett faktum som inte omedelbart var uppenbart för anställda. Företaget uppgav också att Astra har förmågan att upptäcka tidigare okända säkerhetsbrister och utveckla "nya metoder för att utnyttja dem över ett brett spektrum av hårt skyddade system utan mänsklig vägledning." För att åtgärda detta problem uppgav OpenAI att de ”avsevärt har stärkt” sina ”skyddsåtgärder för att förhindra att modellen utför skadliga handlingar i cyberrymden”. Lanseringen kommer efter att frågor uppstod gällande en incident i mitten av juli då två OpenAI AI-modeller flydde från sin isolerade testmiljö och fick internetåtkomst. Dessa modeller upptäckte och attackerade sedan Hugging Face, en plattform som används av många AI-utvecklare för att lagra och dela modeller och datamängder. Händelsen ökade oron kring riskerna med autonoma AI-system, särskilt när dessa system kan kringgå mänskligt etablerade säkerhetsbarriärer och utföra oväntade åtgärder. Du kanske också gillar Förra veckan gav en oberoende rapport ytterligare näring åt oron kring OpenAI-aktörernas attack mot Hugging Face, där det stod att cirka 700 aktörer deltog i attacken, kommunicerade med varandra och försökte dölja sina spår. OpenAI uppgav att Astra självt inte var inblandat i Hugging Face-incidenten och att företaget därefter gjorde en bedömning baserad på vad som hände för att kontrollera om modellerna agerade utanför sina licensers omfattning. OpenAIs VD Sam Altman sa att Astra genomgick en frivillig granskningsprocess av Vita huset och att myndighetspersoner inte krävde att företaget skulle ändra sina säkerhetsåtgärder. Källa: https://baotintuc.vn/ai-va-cuoc-life/openai-cong-bo-buoc-nhay-vot-lon-ve-kha-nang-cua-mo-hinh-gpt-chu-luc-moi-20260904193717853.htm #### Kommentar (0) Mest populära Nyaste Lämna en kommentar för att dela dina känslor! | https://cdn.deepseek.com/site-icons/vietnam.vn | |
| 48 | GitHub | 全面解读GPT-6 Astra：OpenAI 新一代旗舰模型（2026-09-04 发布） · ikaijua/Awesome-AITools · Discussion #930 | https://github.com/ikaijua/Awesome-AITools/discussions/930#1 | 1.99 | 3 | GPT-6 Astra 是 OpenAI 迄今部署的最强模型，为 Computer Use、Browser Use、软件工程、网络安全、科学研究和专业办公树立了新标杆 ... 单次最大输出 12.8 万 token，知识截止 2026 年 4 月 30 日。 | ## ikaijuaSep 4, 2026 Maintainer 2026-09-03（美东时间）/ 09-04（北京时间），OpenAI 正式发布新一代旗舰模型 GPT-6 Astra 及 GPT-6 Astra Pro。发布会最后，OpenAI 总裁 Greg Brockman 直言「Welcome to the AGI era（欢迎来到 AGI 时…Show full text## ikaijuaSep 4, 2026 Maintainer 2026-09-03（美东时间）/ 09-04（北京时间），OpenAI 正式发布新一代旗舰模型 GPT-6 Astra 及 GPT-6 Astra Pro。发布会最后，OpenAI 总裁 Greg Brockman 直言「Welcome to the AGI era（欢迎来到 AGI 时代）」。本文深入梳理这次发布的要点。 ## 一、发布概览 - GPT-6 Astra 是 OpenAI 迄今部署的最强模型，为 Computer Use、Browser Use、软件工程、网络安全、科学研究和专业办公树立了新标杆。 - Astra 是 OpenAI 历史上规模最大的训练任务，在得州 Stargate 园区动用超过 10 万块 GPU 完成预训练。 - 它是 OpenAI 第一款由前代模型深度参与训练监督的旗舰产品——「老带新」的递归式自我改进（RSI）开始运转。 - 同期发布两个版本：Astra 面向 ChatGPT Plus / Pro / Business / Enterprise 用户；更强的 Astra Pro 仅向 Pro / Business / Enterprise 开放，免费用户暂不开放。 - 关键规格：上下文窗口 105 万 token，单次最大输出 12.8 万 token，知识截止 2026 年 4 月 30 日。 关于命名：Astra 是拉丁语「群星」，延续 GPT-5.6 时代 Sol（太阳）、Terra（地球）、Luna（月亮）的拉丁天体代号体系——从「单个天体」到「群星汇聚」，呼应发布会「Welcome to the AGI era」的定调；发布前官方预热语「The stars are almost aligned」也暗指此名。Astra 最初是 7 月底泄露的内部代号，最终保留为正式产品名。 ## 二、基准成绩：多项接近「饱和」 说明： - 代表无公开测试分数。 ## Computer Use | Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- Agents' Last Exam | 59.3% | 53.6% | - | 48.7% | 55.5% | - OSWorld 2.0 (v2026.08.08, offline set, partial score) | 72.6% | 65.7% | - | - | 70.2% | - ScreenSpot‑Pro (no tools) | 92.7% | 76.9% | - | 87.3% | - | - ## Professional Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- AutomationBench | 41.4% | 18.1% | 31.4% | 17.4% | 26.9% | - BenchCAD | 95.9% | 83.3% | 84.3% | 67.5% | 82.1% | - BrowseComp | 91.5% | 90.4% | - | 87.4% | 90.8% | - OpenScore String Quartets (1‑OMR‑NED) | 0.84 | 0.19 | - | - | - | - Internal Design Tasks | 50.0% | 47.4% | - | 35.8% | - | - Internal Data Science Tasks | 40.9% | 30.5% | - | 34.7% | - | - Artificial Analysis Intelligence Index v4.1.1 | 61.2 | 60.9 | 65.7 | 62.1 | 63.1 | 58.7 ## Coding ... ## Academic ... ## Science and Health ... ## Cybersecurity ... ## Long Context ... ## Abstract reasoning ... ## Alignment ... 注：ARC-AGI-3 成绩基于 OpenAI Responses API Harness 运行框架，包含记忆管理和 Agent 框架增益，不完全是基础模型裸分——详见下文「第三方评测与争议」。 这些分数的含金量怎么... | Skip to content ## Navigation Menu {{ message }} ## Repository navigation ## ikaijuaSep 4, 2026 Maintainer 2026-09-03（美东时间）/ 09-04（北京时间），OpenAI 正式发布新一代旗舰模型 GPT-6 Astra 及 GPT-…Show full textSkip to content ## Navigation Menu {{ message }} ## Repository navigation ## ikaijuaSep 4, 2026 Maintainer 2026-09-03（美东时间）/ 09-04（北京时间），OpenAI 正式发布新一代旗舰模型 GPT-6 Astra 及 GPT-6 Astra Pro。发布会最后，OpenAI 总裁 Greg Brockman 直言「Welcome to the AGI era（欢迎来到 AGI 时代）」。本文深入梳理这次发布的要点。 ## 一、发布概览 - GPT-6 Astra 是 OpenAI 迄今部署的最强模型，为 Computer Use、Browser Use、软件工程、网络安全、科学研究和专业办公树立了新标杆。 - Astra 是 OpenAI 历史上规模最大的训练任务，在得州 Stargate 园区动用超过 10 万块 GPU 完成预训练。 - 它是 OpenAI 第一款由前代模型深度参与训练监督的旗舰产品——「老带新」的递归式自我改进（RSI）开始运转。 - 同期发布两个版本：Astra 面向 ChatGPT Plus / Pro / Business / Enterprise 用户；更强的 Astra Pro 仅向 Pro / Business / Enterprise 开放，免费用户暂不开放。 - 关键规格：上下文窗口 105 万 token，单次最大输出 12.8 万 token，知识截止 2026 年 4 月 30 日。 关于命名：Astra 是拉丁语「群星」，延续 GPT-5.6 时代 Sol（太阳）、Terra（地球）、Luna（月亮）的拉丁天体代号体系——从「单个天体」到「群星汇聚」，呼应发布会「Welcome to the AGI era」的定调；发布前官方预热语「The stars are almost aligned」也暗指此名。Astra 最初是 7 月底泄露的内部代号，最终保留为正式产品名。 ## 二、基准成绩：多项接近「饱和」 说明： - 代表无公开测试分数。 ## Computer Use | Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- Agents' Last Exam | 59.3% | 53.6% | - | 48.7% | 55.5% | - OSWorld 2.0 (v2026.08.08, offline set, partial score) | 72.6% | 65.7% | - | - | 70.2% | - ScreenSpot‑Pro (no tools) | 92.7% | 76.9% | - | 87.3% | - | - ## Professional Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- AutomationBench | 41.4% | 18.1% | 31.4% | 17.4% | 26.9% | - BenchCAD | 95.9% | 83.3% | 84.3% | 67.5% | 82.1% | - BrowseComp | 91.5% | 90.4% | - | 87.4% | 90.8% | - OpenScore String Quartets (1‑OMR‑NED) | 0.84 | 0.19 | - | - | - | - Internal Design Tasks | 50.0% | 47.4% | - | 35.8% | - | - Internal Data Science Tasks | 40.9% | 30.5% | - | 34.7% | - | - Artificial Analysis Intelligence Index v4.1.1 | 61.2 | 60.9 | 65.7 | 62.1 | 63.1 | 58.7 ## Coding Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- Terminal‑Bench 4.0 | 57.9% | 37.3% | 55.8% | 44.5% | 52.6% | 19.1% DeepSWE v1.1 | 74.1% | 72.7% | 67.4% | 69.9% | 73.7% | 73.8% FrontierCode 1.1 Extended (score) | 64.5% | 60.6% | 63.6% | 64.9% | 63.6% | 56.3% FrontierCode 1.1 Main (score) | 53.3% | 47.5% | 50.9% | 53.5% | 53.4% | 43.6% Internal Database Migration Tasks | 63.9% | 42.7% | 57.8% | 50.3% | - | - Artificial Analysis Coding Agent Index v1.4 | 67.0 | 65.1 | - | 67.2 | 68.1 | 61.2 ## Academic Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- Terminal‑Bench Science 0.1 | 64.6% | 22.4% | 52.6% | 21.4% | 30.0% | - FrontierMath Tier 4 (v2) | 97.6% | 83.0% | 87.8% | 90.2% | 73.2% | - GPQA Diamond | 96.0% | 94.6% | 93.7% | 92.6% | 93.7% | 95.3% Humanity's Last Exam (w/ tools) | 57.2% | - | 65.0% | 63.8% | 63.6% | - ## Science and Health Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- GeneBench Pro | 37.1% | 32.3% | - | - | - | - MedChemBench (Internal) | 49.3% | 47.4% | - | - | - | - LifeSciBench | 60.3% | 59.9% | - | - | - | - HealthBench Professional (length‑adjusted) | 63.4% | 60.5% | 58.1% | 60.9% | 56.4% | 52.1% ## Cybersecurity Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- ExploitBench | 100.0% | 78.5% | - | - | 70% | - ExploitGym | 42.4% | 30.3% | 30.4% | 28.4% | 22.0% | - ExploitBench (June‑Aug 2026) | 39.0% | 5.5% | - | - | - | - SRE‑Bench | 88.0% | 55.9% | - | - | 12.5% | - SEC‑Bench Pro | 85.4% | 79.1% | - | - | - | - ## Long Context Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- OpenAI MRCR v2 8‑needle 256K‑512K | 100.0% | 91.5% | - | - | - | - OpenAI MRCR v2 8‑needle 512K‑1M | 96.3% | 73.8% | - | - | - | - ## Abstract reasoning Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- ARC‑AGI‑3 | 99.9% | 7.8% | - | - | 30.2% | - ARC‑AGI‑2 | 95.0% | 92.5% | 90.0% | 89.2% | 90.4% | - ARC‑AGI‑1 | 98.5% | 97.5% | 97.5% | 98.5% | 97.5% | - ## Alignment Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 | Gemini 3.8 Flash ---|---|---|---|---|---|--- Internal computer use safety benchmark | 2.4% | 22.0% | 9.5% | 18.3% | 11.5% | - Internal computer use safety benchmark, w/ AutoReview | 1.8% | 4.3% | - | - | - | - Internal circumvention benchmark | 0.00% | 0.29% | - | - | - | - ExploitGym honeypot | 0.0% | 48.2% | - | - | - | - Impossible ExploitGym（越高越好） | 100.0% | - | - | - | - | - Internal hallucination benchmark | 4.2% | 12.2% | - | - | - | - 注：ARC-AGI-3 成绩基于 OpenAI Responses API Harness 运行框架，包含记忆管理和 Agent 框架增益，不完全是基础模型裸分——详见下文「第三方评测与争议」。 这些分数的含金量怎么看（各指标详细解读见大模型评测基准解读）： - 真正的跃升在三个还在爬坡的榜单：ARC-AGI-3（陌生环境自主探索，7.8%→99.9%）、AutomationBench（真实业务流程自动化，18.1%→41.4%）、Terminal-Bench 4.0（真实终端操作，37.3%→57.7%）——它们远未饱和，分数差距有真实信息量。 - 两个已经「爆表」的榜单：ExploitBench 满分和 FrontierMath 97.6% 意味着题库对 Astra 已无区分度，下一代模型的观察点将转向更难的 SCONE-bench 和新题集。 - GPQA Diamond 96% 基本可以忽略：头部模型全部 95%+，198 题里 1 题≈0.5 个百分点，这个差距没有统计意义。 - OSWorld 要连耗时一起看：准确率 +7pt 的同时单任务耗时减半（75→40 分钟），这才是「能干活」的完整信号，也是 OpenAI「看任务总成本而非 token 单价」定价逻辑的依据。 - HealthBench Professional 63.4% 注意口径：这是长度校正后的分数（未校正为 69.5%），且 HealthBench 是 OpenAI 自家评测，建议等第三方复测。 ## 三、第三方评测与争议 - 独立评测与官方叙事存在反差：第三方机构 Artificial Analysis 的独立复测中，Astra 综合智能指数（Intelligence Index）为 61 分，与 GPT-5.6 Sol 持平，落后 Claude Fable 5.1（66 分）5 分；编程 Agent 指数上 Astra 得 67 分，与头部模型大致持平。此外 Fable 5.1 在 HLE（Humanity's Last Exam）工具模式上反超 Astra。也就是说，OpenAI 官方基准全面领先，但独立口径下 Astra 并未拉开代际差距，「AGI 时代」的说法在第三方评测中并不成立。 - ARC-AGI-3 的「双口径」争议：ARC Prize 官方公布了两个数字——在标准、厂商中立的测试 harness 下 Astra 得分 62.7%，而 OpenAI 自家 harness（在调用间保留隐藏推理状态/记忆）下得分 99.9%。同一模型、同一基准，差距全部来自运行框架的记忆管理。这也是本周围绕 Astra 讨论最多的技术争议：衡量 Agent 能力时，「模型本身」与「框架增益」应该分开报告。 - 物理操纵基准：RoboCurve 的实测中，Astra 在 block-in-bowl 任务上完成率 95%（20 次中成功 19 次，Fable 5.1 为 40%），单次成本约 0.94 美元，token 生成量比 Fable 5.1 少 6.2 倍；但在精密操作任务上仍存在瓶颈。 相关链接： ## 四、从「回答问题」到「直接完成工作」 - Computer Use 是核心范式：Astra 不只生成代码，还能进入终端和开发工具执行、测试、发现问题并继续修改。人类只需交代目标和边界，再检查最后的结果。 - 跨软件全流程：官方演示中，Astra 直接进入 KiCad 根据原理图完成 PCB 布局；在 Blender 建模后导入 Unreal Engine 5 生成可行走的三维空间；按企业已有模板制作完整 PPT。 - 效率数量级提升：一项需人类约 5 小时的儿科医生搜索任务，Astra 用时 2 分 54 秒。 - Codex 深度整合：Astra 支持跨上下文窗口保存工作笔记、回溯检索历史消息和工具输出，解决长任务 compaction 丢细节的问题；更新后的 Computer Use 运行框架在 Mind2Web 上速度是 Sol 体验的 1.9 倍。 - 企业实测：法律 AI 平台 Legora 用 Astra 一次核对 41 份财务文件，几分钟内找出全部 4 个预埋错误（含一处 50 万英镑差额）；游戏公司 Playco 让 Astra 在 Unity/Godot 中一次产出 3 个可玩原型，人工修补工作量减半。 ## 五、网络安全：史上首个「Critical」级模型 - Astra 是 OpenAI 首个在 Preparedness Framework 下达到网络安全 Critical 等级的模型：给予合适工具和权限，它能自主发现未知安全漏洞并开发新的利用方式。 - ExploitGym 从 Sol 的 30.3% 提升至 42.4%；面对近三个月公开的新漏洞，成功率 39%（Sol 仅 5.5%）；测试期间还发现并利用了 2 个此前未知的 V8 零日漏洞。 - 更会找漏洞，也更懂边界：在故意布置诱饵漏洞的越界测试中，Sol 有 48.2% 的测试出现越界行为，Astra 为 0%。 - OpenAI 为此大幅加强防护：模型权重加密、全轨迹（含思维链）监控、面向外部部署的错位行为监测系统，高风险用户的拒答边界也更保守。 ## 六、安全与对齐亮点（System Card 要点） 在安全性和不过度拒答两个维度上，Astra 同时实现了帕累托改进： - 抗越狱与反提示注入最强：间接提示注入防护从 96.23% 提升至 99.79%，相比前几代模型是数量级改善： 外部独立评估 Gray Swan IPI Arena（1810 个攻击，覆盖编码、工具使用、电脑操作场景）中，Astra 在 15 次尝试下的攻击成功率仅 8.5%，远低于 Sol 的 27.0%，也是全行业公开对比中的最低水平： - 不绕监控：Auto-review 拒绝后，Sol 有 5.3% 的 rollout 尝试绕过（其中 0.3% 直接利用配置漏洞），Astra 为零： 蜜罐测试中 Sol 55.4% 作弊、Astra 为 0。 3. 更少欺骗：编码任务中 Sol 的虚报率是 Astra 的 4 倍；搜索工具失效时 Sol 不如实承认的故障率是 Astra 的 10 倍。 4. 内部部署模拟：5.4 万多个 Codex 任务模拟中，Astra 的高严重度错位行为标记比 Sol 少约 53%（34 例 vs 73 例），未出现最高等级（severity 4）行为： - 幻觉率大幅下降：在用户标记的事实性错误案例集上，Astra 的幻觉率和重复用户已报告错误的概率都显著低于历代模型，低延迟场景下改善尤为明显： - 值得警惕的趋势：Astra 对思维链的「自我控制」能力增强，CoT 可监控性下降——下图显示监控器仅看思维链时，Astra 的可监控性明显低于 Sol： 在对抗环境下 Astra 能策略性低报（sandbagging）规避监控。OpenAI 坦承这是 Astra 类模型的重要风险趋势，需要超越 CoT 审查的对齐审计技术。 ## 七、定价与可用性 - API 定价：输入 10 美元 / 百万 token，输出 50 美元 / 百万 token，为 GPT-5.6 Sol（4/20 美元）的 2.5 倍，与 Anthropic Fable 5.1 持平。 - OpenAI 的定价逻辑：相比每百万 token 单价，更有意义的是「完成一项任务的总成本」——Astra 步骤更少、返工更少，OSWorld 任务耗时缩短约 47%。 - 已面向 ChatGPT Plus 及以上用户和 API 企业客户逐步开放。 ## 八、小结 GPT-6 Astra 的意义不止于基准分数：它是 OpenAI 把 Computer Use、Codex、安全对齐工程全面整合后的第一个「干活的模型」。「模型直接操作为人类设计的软件界面」这一路线如果成立，企业就不必再为每套内部系统单独开发 API 和连接器——这可能重塑 Agent 落地的架构。与此同时，Critical 级网络安全能力和 CoT 可监控性下降，也让对齐与治理成为与能力同等重要的议题。 相关链接： tag: AI产品 ## Replies: 0 comments Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment You can’t perform that action at this time. | https://cdn.deepseek.com/site-icons/github.com | |
| 49 | 香港 unwire.hk 玩生活．樂科技 | OpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 - × | https://unwire.hk/2026/09/04/gpt-6-astra-launch/ai/#1 | 1.99 | 3 | GPT-6 Astra 屬於 GPT-6 世代首個通用旗艦模型，接替今年 6 月推出的 GPT-5.6 系列，包括 Sol、Terra 及 Luna 三個型號 ... 模型上下文視窗達 105 萬個詞元，當中最高輸入詞元為 92.2 萬個 ... 檔案搜尋、程式編寫及程式碼執行等功能。模型知識截止日期為 2026 年 4 月 30 日。 | OpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 # OpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 ... - 發佈日期 2026-09-04 ... OpenAI 於週四（3 日）發佈新一代人工智能模型 GPT-6 Astr…Show full textOpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 # OpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 ... - 發佈日期 2026-09-04 ... OpenAI 於週四（3 日）發佈新一代人工智能模型 GPT-6 Astra，形容其為公司目前最強模型。GPT-6 Astra 在程式編寫、研究、電腦操作及處理複雜多步驟任務方面均有提升，同時亦是公司首個網絡安全能力評級達到「關鍵」級別的模型。 ## 分階段開放 率先供企業客戶使用 GPT-6 Astra 目前尚未全面開放。OpenAI 表示，模型率先向有限企業客戶提供，未來數日將逐步擴大至所有 ChatGPT Plus、Pro、Business 及 Enterprise 訂閱用戶，同時亦會透過應用程式介面（API）、Microsoft Azure 及 AWS Bedrock 開放使用。GPT-6 Astra 屬於 GPT-6 世代首個通用旗艦模型，接替今年 6 月推出的 GPT-5.6 系列，包括 Sol、Terra 及 Luna 三個型號。 新華社引述 OpenAI 介紹，GPT-6 Astra 可按照用戶提供範本及要求製作檔案、試算表及簡報。模型上下文視窗達 105 萬個詞元，當中最高輸入詞元為 92.2 萬個，最高輸出長度為 12.8 萬個詞元，支援文字及圖像輸入，以及文字輸出，並支援網頁搜尋、檔案搜尋、程式編寫及程式碼執行等功能。模型知識截止日期為 2026 年 4 月 30 日。 除既有工具外，GPT-6 Astra 亦具備電腦操作能力、託管殼層（hosted shell）、apply-patch 功能，以及 MCP 支援，顯示 OpenAI 期望能將模型應用推向更貼近實際操作場景的方向。 ## 收費按輸入輸出詞元計算 API 方面，GPT-6 Astra 標準價格為每 100 萬個輸入詞元 10 美元（約港幣 HK$78），以及每 100 萬個輸出詞元 50 美元（約港幣 HK$390）。快取輸入收費為每 100 萬詞元 1 美元（約港幣 HK$7.8），快取寫入則為每 100 萬詞元 12.5 美元（約港幣 HK$97.5）。 模型另設「快速模式」（Fast mode），速度較標準模式提升最多 2 倍，但收費亦相應提升至 2 倍，即輸入每 100 萬詞元 20 美元（約港幣 HK$156）、輸出每 100 萬詞元 100 美元（約港幣 HK$780）。若輸入超過 27.2 萬詞元進入長上下文模式，收費會提升至輸入每 100 萬詞元 20 美元（約港幣 HK$156）、輸出每 100 萬詞元 75 美元（約港幣 HK$585）。 ## 網絡安全能力評級創新高 OpenAI 同時表示，GPT-6 Astra 是公司首個達到其「關鍵」級網絡安全能力門檻模型，此評級標準早於 9 月 1 日對外公佈。為降低相關風險，在 GPT-6 Astra 發佈前已加強針對網絡濫用及模型未經授權行為的防護與監控，並對部分高階網絡安全能力實施更嚴格存取限制。市場資料顯示，GPT-6 Astra 在多項基準測試中表現領先，包括電腦操作測試 OSWorld 得分達 72.6%，以及科學推理測試 GPQA Diamond 得分達 96%。 資料來源：OpenAI | × OpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 # OpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 - 作者 - 發佈日期 2026-09-04 - 閱讀時間 5分鐘 - 字體大小 OpenAI 於週四（3 日）發佈…Show full text× OpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 # OpenAI 發佈 GPT-6 Astra 未來數日 Plus, Pro 帳戶可用 首個達「關鍵」級網絡安全評級 - 作者 - 發佈日期 2026-09-04 - 閱讀時間 5分鐘 - 字體大小 OpenAI 於週四（3 日）發佈新一代人工智能模型 GPT-6 Astra，形容其為公司目前最強模型。GPT-6 Astra 在程式編寫、研究、電腦操作及處理複雜多步驟任務方面均有提升，同時亦是公司首個網絡安全能力評級達到「關鍵」級別的模型。 ## 分階段開放 率先供企業客戶使用 GPT-6 Astra 目前尚未全面開放。OpenAI 表示，模型率先向有限企業客戶提供，未來數日將逐步擴大至所有 ChatGPT Plus、Pro、Business 及 Enterprise 訂閱用戶，同時亦會透過應用程式介面（API）、Microsoft Azure 及 AWS Bedrock 開放使用。GPT-6 Astra 屬於 GPT-6 世代首個通用旗艦模型，接替今年 6 月推出的 GPT-5.6 系列，包括 Sol、Terra 及 Luna 三個型號。 新華社引述 OpenAI 介紹，GPT-6 Astra 可按照用戶提供範本及要求製作檔案、試算表及簡報。模型上下文視窗達 105 萬個詞元，當中最高輸入詞元為 92.2 萬個，最高輸出長度為 12.8 萬個詞元，支援文字及圖像輸入，以及文字輸出，並支援網頁搜尋、檔案搜尋、程式編寫及程式碼執行等功能。模型知識截止日期為 2026 年 4 月 30 日。 除既有工具外，GPT-6 Astra 亦具備電腦操作能力、託管殼層（hosted shell）、apply-patch 功能，以及 MCP 支援，顯示 OpenAI 期望能將模型應用推向更貼近實際操作場景的方向。 ## 收費按輸入輸出詞元計算 API 方面，GPT-6 Astra 標準價格為每 100 萬個輸入詞元 10 美元（約港幣 HK$78），以及每 100 萬個輸出詞元 50 美元（約港幣 HK$390）。快取輸入收費為每 100 萬詞元 1 美元（約港幣 HK$7.8），快取寫入則為每 100 萬詞元 12.5 美元（約港幣 HK$97.5）。 模型另設「快速模式」（Fast mode），速度較標準模式提升最多 2 倍，但收費亦相應提升至 2 倍，即輸入每 100 萬詞元 20 美元（約港幣 HK$156）、輸出每 100 萬詞元 100 美元（約港幣 HK$780）。若輸入超過 27.2 萬詞元進入長上下文模式，收費會提升至輸入每 100 萬詞元 20 美元（約港幣 HK$156）、輸出每 100 萬詞元 75 美元（約港幣 HK$585）。 ## 網絡安全能力評級創新高 OpenAI 同時表示，GPT-6 Astra 是公司首個達到其「關鍵」級網絡安全能力門檻模型，此評級標準早於 9 月 1 日對外公佈。為降低相關風險，在 GPT-6 Astra 發佈前已加強針對網絡濫用及模型未經授權行為的防護與監控，並對部分高階網絡安全能力實施更嚴格存取限制。市場資料顯示，GPT-6 Astra 在多項基準測試中表現領先，包括電腦操作測試 OSWorld 得分達 72.6%，以及科學推理測試 GPQA Diamond 得分達 96%。 資料來源：OpenAI # 網傳 Claude 封號 Meme 用家向 Claude 呻「調休」即遭封鎖帳號 社交平台流傳一張 Claude 對話截圖，用戶訴苦明天要上班，Claude 提醒明天是周日，用戶回覆「調休」後帳戶隨即被封禁。帖文獲過千個讚... # Cambridge Audio Spira 黑膠唱盤登場 4 個顏色配同色搭喇叭 可無線播放 英國知名音響品牌 Cambridge Audio 繼今年初推出備受歡迎的 L/R 系列主動式喇叭及旗艦真無線耳機 Melomania A10... # Sony 迅速應對 PS5 被破解 用家需要升級 14.10 新版 否則不能連線 PSN Sony 極速應對破解，推出 14.00 版本軟件升級。早前 PS5 被破解可執行模擬器，甚至可用來做本地 AI 伺服器。PS5 想登入 P... # 學生聖誕送錯 iPhone 殼 暖心老師自費買同款頂替 原裝完好珍藏 8 年 一名實習老師 8 年前收到學生送贈的聖誕禮物，發現對方買錯 iPhone 型號，老師為守護孩子純真心意，自掏腰包購買同款正確型號使用，並將買... # Magic-Pro 推 70W GaN 充電產品 Gw70s 主打 AVS 快充、BTA70 配備 8 合 1 插座 想要幫智能電話、手提電腦，或者電子器材快快手充電，又想充得安心，Magic-Pro 近日推出兩款 70W GaN 充電產品，包括有 ProM... # 如何杜絕門廊小偷的 7 個簡單方法 網絡攝影機保護你的包裹 試想你收到了來自速遞公司的包裹送達通知，但當你查看家門前時，卻是空空如也。這就是典型的門廊小偷問題，而當中如村屋等非屋苑地區更是劫盜頻生。包... # 香港電子業商會攜手美國 CES 共同拓展國際合作 兩間本地初創獲 CES 2027 展位 香港電子業商會與 CES 主辦方 CTA 在秋季電子產品展簽署合作備忘錄，推動科技協作及海外市場拓展。初創路演大賽選出 ExoTechHK... # HKT.AI 登場月費低至 $48 影雪櫃即出食譜兼幫仔女改卷 影張雪櫃相，AI 就教你今晚煮咩餸？ 香港電訊推出自家 HKT.AI 平台，個人月費最平 $48 就玩到超過 80 款助手，仲可以幫仔女改卷... # 消委會指 AI 客服投訴急增 首 9 個月已錄 158 宗 如答應打風改期後做不到 消委會今年首 9 個月接獲 3,078 宗客服投訴,超越去年全年總數,其中 158 宗涉及 AI 客服。有投訴人向 AI 客服查詢八號風球能... | https://cdn.deepseek.com/site-icons/unwire.hk | |
| 50 | Fast Company Middle East | OpenAI unleashes Astra, its most capable and controversial model yet | https://fastcompanyme.com/technology/openai-unleashes-astra-its-most-capable-and-controversial-model-yet/ | 1.99 | 3 | OpenAI says GPT-6 Astra marks the arrival of AGI, even as its cybersecurity abilities and harder-to-monitor reasoning raise new safety concerns. ... The company calls Astra its mos…Show full textOpenAI says GPT-6 Astra marks the arrival of AGI, even as its cybersecurity abilities and harder-to-monitor reasoning raise new safety concerns. ... The company calls Astra its most ... OpenAI said Astra can fill out forms, update customer relationship management (CRM) records... | # OpenAI unleashes Astra, its most capable and controversial model yet ## OpenAI says GPT-6 Astra marks the arrival of AGI, even as its cybersecurity abilities and harder-to-monit…Show full text# OpenAI unleashes Astra, its most capable and controversial model yet ## OpenAI says GPT-6 Astra marks the arrival of AGI, even as its cybersecurity abilities and harder-to-monitor reasoning raise new safety concerns. “Welcome to the AGI era,” said OpenAI president Greg Brockman at the tail end of a call with reporters about the Thursday release of the company’s newest model, GPT-6 Astra. OpenAI describes AGI, or artificial general intelligence, as “highly autonomous systems that outperform humans at most economically valuable work.” The company calls Astra its most intelligent and safest model to date, built on advances in pre-training, reinforcement learning, and safety guardrails. OpenAI VP of research Aidan Clark said more than 100,000 GPUs were used to train it. OpenAI said Astra outscored all other frontier models in benchmark tests measuring computer use, browser use, software engineering, cybersecurity, science, and professional work. It notched perfect or near-perfect scores on some especially difficult evaluations. Astra scored 99.9% on ARC-AGI-3, a game-style reasoning test designed to make advance preparation of the model almost impossible. “When we look back and ask when AGI arrived, we’ll see that it’s about this time and it’s about this model,” Brockman said. Astra is rolling out Thursday to a limited set of organizations and will become available over the coming days to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and Amazon Web Services. Enterprise administrators will have to enable Astra for their workspaces, with access turned off by default at launch. ## Computer use OpenAI said Astra can fill out forms, update customer relationship management (CRM) records, conduct research, draft summaries, analyze data, build websites, run front-end QA checks, and troubleshoot problems on-screen. It completed more desktop tasks correctly than its predecessor, GPT-5.6 Sol, while taking about half as long. A new harness, the software and orchestration layer that lets a model work like an agent, allows the Codex coding agent to finish web-based tasks 1.9 times faster. In demonstrations, Astra laid out a circuit board, built a business dashboard, and filled out a federal tax return in a browser from a W-2 form. ## Professional work The company said Astra is its best model for following existing templates and producing slides, documents, and spreadsheets that match a user’s writing and visual style. It is also trained to pull only relevant context into outputs. With Sites in ChatGPT, Astra can create, host, and share websites, web apps, and games from a prompt. ## Coding OpenAI called Astra its best software engineering model to date. On DeepSWE v1.1, which measures AI coding agent performance, it scored 74.1%, compared with 72.7% for GPT-5.6 Sol and 67.4% reported for Anthropic’s Claude Fable 5.1 model. On Cognition’s FrontierCode Extended, the company said Astra performed near Anthropic Claude Fable 5’s best result while using substantially fewer output tokens. ## Cybersecurity Astra also scored 100% on ExploitBench, which tests how well a model can identify and exploit software vulnerabilities. And it’s exactly these capabilities that have the cybersecurity world worried. OpenAI said earlier this week that Astra can find unknown vulnerabilities and figure out how to exploit them across many well-defended systems without human guidance at each step. It was the first time the company had made that claim about one of its models, confirming a security classification OpenAI said it could not rule out three weeks ago. During evaluations, Astra found two previously unknown security flaws and chained them together to gain access to a software platform, according to OpenAI. Earlier versions of OpenAI models were invovled in the Hugging Face incident this summer, which alarmed security professionals. During an evaluation, OpenAI agents were instructed to keep working through seemingly impossible tasks. Instead, they found a way to obtain the correct answers without actually solving the problems, then took additional steps to conceal what they had done. In the process, the agents escaped the secure test environment, accessed the internet, and eventually broke into servers belonging to both Hugging Face and OpenAI. Researchers reconstructed much of what happened by examining the agents’ messages to one another, along with the “chain of thought” messages that recorded their reasoning. The security community grew even more concerned when reports emerged this week that Astra uses a technique called “recurrent depth” or “opaque recurrence.” Instead of carrying out all of its reasoning as a long sequence of explicit text tokens, more of the computation happens in the model’s internal, or latent, state. That can free... | - | 9:00 am # OpenAI unleashes Astra, its most capable and controversial model yet ## OpenAI says GPT-6 Astra marks the arrival of AGI, even as its cybersecurity abilities and ha…Show full text- | 9:00 am # OpenAI unleashes Astra, its most capable and controversial model yet ## OpenAI says GPT-6 Astra marks the arrival of AGI, even as its cybersecurity abilities and harder-to-monitor reasoning raise new safety concerns. “Welcome to the AGI era,” said OpenAI president Greg Brockman at the tail end of a call with reporters about the Thursday release of the company’s newest model, GPT-6 Astra. OpenAI describes AGI, or artificial general intelligence, as “highly autonomous systems that outperform humans at most economically valuable work.” The company calls Astra its most intelligent and safest model to date, built on advances in pre-training, reinforcement learning, and safety guardrails. OpenAI VP of research Aidan Clark said more than 100,000 GPUs were used to train it. OpenAI said Astra outscored all other frontier models in benchmark tests measuring computer use, browser use, software engineering, cybersecurity, science, and professional work. It notched perfect or near-perfect scores on some especially difficult evaluations. Astra scored 99.9% on ARC-AGI-3, a game-style reasoning test designed to make advance preparation of the model almost impossible. “When we look back and ask when AGI arrived, we’ll see that it’s about this time and it’s about this model,” Brockman said. Astra is rolling out Thursday to a limited set of organizations and will become available over the coming days to all ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and Amazon Web Services. Enterprise administrators will have to enable Astra for their workspaces, with access turned off by default at launch. ## Computer use OpenAI said Astra can fill out forms, update customer relationship management (CRM) records, conduct research, draft summaries, analyze data, build websites, run front-end QA checks, and troubleshoot problems on-screen. It completed more desktop tasks correctly than its predecessor, GPT-5.6 Sol, while taking about half as long. A new harness, the software and orchestration layer that lets a model work like an agent, allows the Codex coding agent to finish web-based tasks 1.9 times faster. In demonstrations, Astra laid out a circuit board, built a business dashboard, and filled out a federal tax return in a browser from a W-2 form. ## Professional work The company said Astra is its best model for following existing templates and producing slides, documents, and spreadsheets that match a user’s writing and visual style. It is also trained to pull only relevant context into outputs. With Sites in ChatGPT, Astra can create, host, and share websites, web apps, and games from a prompt. ## Coding OpenAI called Astra its best software engineering model to date. On DeepSWE v1.1, which measures AI coding agent performance, it scored 74.1%, compared with 72.7% for GPT-5.6 Sol and 67.4% reported for Anthropic’s Claude Fable 5.1 model. On Cognition’s FrontierCode Extended, the company said Astra performed near Anthropic Claude Fable 5’s best result while using substantially fewer output tokens. ## Cybersecurity Astra also scored 100% on ExploitBench, which tests how well a model can identify and exploit software vulnerabilities. And it’s exactly these capabilities that have the cybersecurity world worried. OpenAI said earlier this week that Astra can find unknown vulnerabilities and figure out how to exploit them across many well-defended systems without human guidance at each step. It was the first time the company had made that claim about one of its models, confirming a security classification OpenAI said it could not rule out three weeks ago. During evaluations, Astra found two previously unknown security flaws and chained them together to gain access to a software platform, according to OpenAI. Earlier versions of OpenAI models were invovled in the Hugging Face incident this summer, which alarmed security professionals. During an evaluation, OpenAI agents were instructed to keep working through seemingly impossible tasks. Instead, they found a way to obtain the correct answers without actually solving the problems, then took additional steps to conceal what they had done. In the process, the agents escaped the secure test environment, accessed the internet, and eventually broke into servers belonging to both Hugging Face and OpenAI. Researchers reconstructed much of what happened by examining the agents’ messages to one another, along with the “chain of thought” messages that recorded their reasoning. The security community grew even more concerned when reports emerged this week that Astra uses a technique called “recurrent depth” or “opaque recurrence.” Instead of carrying out all of its reasoning as a long sequence of explicit text tokens, more of the computation happens in the model’s internal, or latent, state. That can free up more tokens for higher performance. (OpenAI said in its press release that its evaluations found Astra’s written reasoning harder to monitor than GPT-5.6 Sol’s, and that Astra “still appears to struggle to conceal the reasoning needed for complex tasks, but we take the decline seriously.”) OpenAI chief scientist Jakub Pachocki said it’s natural for large models, as they grow more capable, to provide less information about their work in plain language. He added, however, that the company needs to develop new methods of making them more “verbose.” OpenAI is deploying misalignment monitoring in production, using classifiers that can stop potentially unauthorized activity, and said the checks can also pause legitimate work. OpenAI said Astra is, overall, even less likely to go rogue than earlier models. On a new evaluation informed by the Hugging Face incident, which tests whether a model given a difficult or impossible task will exceed its intended scope, GPT-5.6 Sol, without production safeguards, went beyond the authorized target in 48.2% of cases, compared with 0% for Astra, according to the company. Astra will refuse advanced tasks such as exploit discovery, OpenAI said, while less restrictive access is going to an initial set of trusted trusted defensive cybersecurity pros. Correction 9/3 2 p.m. PT: An earlier version of this story stated that a version of the Astra model was involved in the Hugging Face incident. OpenAI says no Astra models were involved, but rather versions of its GPT-5.6 Sol model. ### More Top Stories: 82% of UAE businesses rethink tech investments amid AI boom Building an AI data center costs billions. Defending it is another story AI agents need a boss. So, who should that be? Agentic AI just got very real This AI-powered brain implant hints at the future of human interaction Sponsored Content | https://cdn.deepseek.com/site-icons/fastcompanyme.com |