GPT-6 Astra

0:04 GPT-6 Astra is OpenAI's flagship model, released on Sep 3, 2026 and described by OpenAI as "Our most capable model for the most demanding work." In the API it costs $10 per million input tokens and $50 per million output tokens, five times GPT-6.1 Sol. In ChatGPT it runs the Pro reasoning option (GPT-6 Pro) in the Chat tab and is selectable in ChatGPT Work and Codex. It is also the first model OpenAI rates Critical for cybersecurity capability, which shaped how it was released.

Quick facts

Specs from the GPT-6 Astra model page, OpenAI API docs, read Nov 6, 2026.

Who can use GPT-6 Astra in ChatGPT?

In the Chat tab, Astra is behind the Pro option: "GPT-6 Pro, powered by GPT-6 Astra, is available in ChatGPT for Pro $100, Pro $200, Business, and Enterprise plans". That help page does not name Pro $500 in the sentence, though the pricing page lists "Pro reasoning powered by GPT-6 Astra" for every Pro tier. GPT-6 Pro has a weekly limit even on Pro; OpenAI publishes no number for it. On Enterprise and Edu, "GPT-6 Astra is off by default at launch" until an admin turns it on.

In ChatGPT Work and Codex, Pro $100, Pro $200 and Business Premium seats can spend their whole allowance on Astra, while Plus and Business Standard include limited Astra use. OpenAI estimates a Plus or Business Standard user gets 5 to 45 local Astra messages per five hours, against 15 to 160 on GPT-6.1 Sol; Pro has no five-hour limit. Free and Go plans do not get Astra. In Codex it needs CLI version 0.153.0 or newer.

Ultrafast and Fast mode

0:00 Astra can run faster for more money. Fast mode costs twice the Standard rate ($20 / $100 per 1M tokens) and, unlike Fast on other models, has no latency guarantee for Astra. Ultrafast, added for Astra in the API on Sep 29, 2026, costs six times Standard: $60 input and $300 output per 1M tokens. At DevDay OpenAI said Ultrafast runs up to 8x faster in Codex (about 300 tokens per second) and up to 6x in the API. OpenAI's own Ultrafast video gives a higher API figure: tokens over seven times faster than the Standard tier and over four times faster than Fast. In ChatGPT, Astra Ultrafast is only on Pro 500, at $500 a month, and eligible Enterprise and Edu, where it uses included usage at 8 times the Standard rate.

Benchmarks

OpenAI's announcement compares Astra with GPT-5.6 Sol and rival models, taking each model's best score at any effort. These are OpenAI's own figures:

Source: Introducing GPT-6 Astra, OpenAI, Sep 3, 2026. OpenAI published no SWE-bench Verified or SWE-bench Pro score for Astra.

Independent, dated measurements are less flattering. On the Artificial Analysis Intelligence Index (v5), Astra (max) scores 57, behind Gemini 4 Carbon (high) at 60 and Claude Opus 5.5 (max) at 59, and it costs about $6.24 per index task against $10.04 for Claude Opus 5.5. OpenAI's own table had cited an older index version (v4.1.1), where Astra scored 61.2; Fireship's Sep 4, 2026 Code Report noted that this launch-day score matched GPT-5.6 Sol's and trailed Claude Fable 5.1 by five points. On LMArena's text board (Oct 27, 2026), Astra (Max) is 37th; on its WebDev board it is 2nd, behind Claude Opus 5.5. Anthropic's Sep 22, 2026 table puts Claude Opus 5.5 ahead of Astra on Terminal-Bench 4.0 (66.4% to 57.9%), while Astra keeps the lead on AutomationBench. See OpenAI Benchmarks.

Safety rating and release

OpenAI's system card calls Astra "our first model to reach the Critical level of cybersecurity capability", rates it High (not Critical) for biological and chemical risk, and below High for AI self-improvement. On misaligned actions in realistic work tasks, the card reports Astra at 3.4% against 18.8% for GPT-5.6 Sol. CNBC reported that Astra went through a formal review with the Trump administration before release, and that companies in OpenAI's Daybreak cybersecurity program got it first. OpenAI's announcement says it reached everyone else on Plus, Pro, Business and Enterprise, plus the API, Azure and AWS Bedrock, "over the coming days".

When should you use GPT-6 Astra?

0:28 Use Astra for the hardest, longest tasks where a wrong answer costs more than the tokens: agentic coding across a large codebase, computer use, math and science problems, and security work. For most other jobs OpenAI itself points elsewhere: GPT-6.1 Sol "nearly matches" Astra on agentic coding, computer use and professional work at one-fifth of the token price, and GPT-6.1 Terra is cheaper still. In ChatGPT, Astra's weekly rationing means everyday questions belong on Instant or Medium.

In the API, Astra supports the Responses API tools (web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search), Chat Completions and the Batch API, where it costs $5 / $25 per 1M tokens. It does not support the Realtime API or Fine-Tuning. A legal variant, gpt-6-astra-law, is listed at the same $10 / $50, on promotional pricing "at least through December 31, 2026". For the cyber and life-science versions, see GPT-Rosalind and Cyber Models.

What testers found

11:34 In a video recorded on Sep 3, 2026, before the public rollout, Matt Wolfe used his early access to let Astra take over his computer. From one request it opened Blender and modeled a wolf figure in about 8 minutes, then added a 50-bone skeleton and, when asked, a running animation. He found the motion a little awkward but the result impressive for someone who, in his words, has "no idea how to use Blender".

7:04 In the same video his usual game test, a 3D clone of Megabonk, came back in 8 minutes, where such builds had taken him an hour and a half to two hours before; he wondered whether the small number of early users made it quicker. He rated it one of the best-looking versions he had gotten, though the knight's hammer was poorly animated.

10:06 In a Sep 13, 2026 video, Lisa Zachmann gave GPT-6 Astra in ChatGPT Work and Claude Fable 5.1 the same one-prompt request for a personal website. Astra's one-page site was done in 8 minutes 32 seconds, and she judged it the better result for its overall presentation and its use of her photos, while noting that both models picked her brand colors.

Frequently asked questions

How much does GPT-6 Astra cost? In the OpenAI API, GPT-6 Astra costs $10 per million input tokens, $1 cached and $50 per million output tokens for prompts up to 272K tokens, as of Nov 6, 2026. In ChatGPT it comes with Pro (from $100 a month), Business and Enterprise in Chat, and in limited amounts with Plus in Work and Codex.

Is GPT-6 Astra available on ChatGPT Plus? Partly. Plus includes limited GPT-6 Astra use in ChatGPT Work and Codex (an estimated 5 to 45 local messages per five hours), but not the GPT-6 Pro option in the Chat tab.

What is GPT-6 Astra's context window? GPT-6 Astra accepts 1,050,000 tokens in the API, up to 922,000 of them input, and writes up to 128,000 tokens. Its knowledge cutoff is Apr 30, 2026.

Is GPT-6 Astra the best AI model? By OpenAI's own tables it leads on many agent, math and cyber tests. On the dated independent leaderboards read on Nov 6, 2026 it is not first: Gemini 4 Carbon and Claude Opus 5.5 rank above it on Artificial Analysis's index, and Claude Opus 5.5 leads LMArena's WebDev board.

What is GPT-6 Astra Ultrafast? Ultrafast is a speed mode for Astra and GPT-6.1 Sol that OpenAI says runs up to 8x faster in Codex. In the API it costs six times the Standard price; in ChatGPT it is only on Pro 500 and eligible Enterprise and Edu.

Sources

• Introducing GPT-6 Astra, OpenAI, Sep 3, 2026
• GPT-6 Astra System Card, OpenAI Deployment Safety Hub, Sep 3, 2026, read Nov 6, 2026
• GPT-6 Astra model page, OpenAI API docs, read Nov 6, 2026
• Models, OpenAI API docs, read Nov 6, 2026
• Pricing, OpenAI API docs, read Nov 6, 2026
• Fast mode, OpenAI API docs, read Nov 6, 2026
• API changelog, OpenAI, read Nov 6, 2026
• GPT-6 and other models in ChatGPT, OpenAI Help Center, read Nov 6, 2026
• ChatGPT Work and Codex, OpenAI Help Center, read Nov 6, 2026
• ChatGPT Enterprise and Edu models and limits, OpenAI Help Center, read Nov 6, 2026
• About ChatGPT Pro tiers, OpenAI Help Center, read Nov 6, 2026
• ChatGPT pricing, OpenAI, read Nov 6, 2026
• Codex pricing, OpenAI, read Nov 6, 2026
• OpenAI announces rollout of GPT-6 Astra model, CNBC, Sep 3, 2026
• OpenAI unveils GPT-6.1 Sol at DevDay, Unite.AI, Sep 29, 2026
• Introducing Claude Opus 5.5, Anthropic, Sep 22, 2026
• Artificial Analysis model leaderboard, read Nov 6, 2026
• LMArena Text leaderboard, updated Oct 27, 2026, read Nov 6, 2026
• LMArena WebDev leaderboard, updated Oct 27, 2026, read Nov 6, 2026
• Introducing GPT-6 Astra for developers, OpenAI, Sep 4, 2026 (video)
• Build faster with Ultrafast, OpenAI, Sep 29, 2026 (video)
• Did OpenAI actually build AGI? GPT-6 Astra first look, Fireship, Sep 4, 2026 (video)
• GPT-6 Astra Is Finally Here (And It’s REALLY Good), Matt Wolfe, Sep 3, 2026 (video)
• ChatGPT vs. Claude: GPT-6 Astra vs. Fable 5.1, Lisa Zachmann, Sep 13, 2026 (video)