GPT-6 Astra: OpenAI's specs, prices, benchmarks, and what happens to your prompts
The Morse teamPublished 10 min read
GPT-6 Astra is the reasoning model OpenAI describes as "for complex coding and professional work," released on September 3, 2026. As of November 2026 it's on the OpenAI API as gpt-6-astra with a 1,050,000-token context window, takes text and images and answers in text, and costs $10 per million input tokens and $50 per million output tokens; input doubles to $20 and output rises to $75 once a prompt passes 272,000 tokens. It's no longer OpenAI's newest Astra. GPT-6.2 Astra arrived on November 17, 2026, priced at $8 and $40.
We read OpenAI's GPT-6 Astra model page, its API changelog, pricing page, model guide, data controls, and misalignment monitoring docs, the launch post, the "Path to Astra" safety update, the system card, ChatGPT's pricing page, and three ChatGPT help articles on November 18, 2026. To OpenAI's credit, its system card keeps a dated change log, including a correction to a third-party biology score, and its own benchmark table shows Claude Fable 5.1 ahead of Astra on two rows.
What is GPT-6 Astra for?
Hard, long tasks. OpenAI's model page lists "complex reasoning, coding, computer use, research, and document creation," and the changelog entry for September 3 says it's built to carry a task "from an initial request to a finished result, using the context and tools you provide." OpenAI's launch post calls its own model "the world's most intelligent and aligned model" and "the best model for software engineering to date."
Its model guide adds a few traits builders run into. Astra is "more likely to ask the user a question when additional input could materially change the result," it "tends toward detailed, formatted responses," and it can be "more sensitive to instructions contained in skills and other files, such as AGENTS.md," which OpenAI "strongly" recommends auditing.
GPT-6 Astra specs at a glance
As of November 18, 2026, from OpenAI's own pages.
| Spec | GPT-6 Astra | Where OpenAI says it |
|---|---|---|
| Released | September 3, 2026 | Changelog |
| API model name | gpt-6-astra (one snapshot) | Model page |
| Context window | 1,050,000 tokens, 922,000 of them input | Model page |
| Longest reply | 128,000 tokens | Model page |
| Knowledge cutoff | April 30, 2026 | Model page |
| Reads and writes | Text and images in, text out | Model page |
| Reasoning effort | Low, medium, high, xhigh, or max; no "none" | Changelog |
| Not supported | Custom temperature or top_p, log probabilities, fine-tuning, Realtime | Model page |
| Tools | Web search, file search, code interpreter, computer use, MCP, and more, on the Responses API only | Model guide |
| Also on | Amazon Bedrock (openai.gpt-6-astra), Google Cloud in private preview | Model page |
What is GPT-6 Astra's context window in ChatGPT?
Smaller than on the API, going by ChatGPT's own plan table. The API takes 1,050,000 tokens. ChatGPT's pricing page lists a "GPT Reasoning total context window" of 256K on Plus and 400K on Pro, without naming a model, so a document that fits in one API call may not fit in one ChatGPT chat.
How much does GPT-6 Astra cost?
As of November 18, 2026, OpenAI's pricing page bills it per million tokens in two context bands. Short context means 272,000 input tokens or fewer.
| Processing tier | Input | Cached input | Output | Input above 272K | Output above 272K |
|---|---|---|---|---|---|
| Standard | $10 | $1 | $50 | $20 | $75 |
| Batch or Flex | $5 | $0.50 | $25 | $10 | $37.50 |
| Fast | $20 | $2 | $100 | $40 | $150 |
| Ultrafast | $60 | $6 | $300 | $120 | $450 |
The step applies to the whole request: the model page says prompts over 272K input tokens are priced "for the full request." Cache writes cost 1.25 times the input rate. Ultrafast, added to the Responses API on September 29, runs with "global processing and US data residency" (OpenAI).
What one request costs, next to GPT-6.2 Astra and GPT-6.1 Sol
By our arithmetic from those Standard rates, each with a 5,000-token reply and no caching:
| Prompt size | GPT-6 Astra | GPT-6.2 Astra | GPT-6.1 Sol |
|---|---|---|---|
| 50,000 tokens | $0.75 | $0.60 | $0.15 |
| 250,000 tokens | $2.75 | $2.20 | $0.55 |
| 300,000 tokens | $6.38 | $5.10 | $1.28 |
Going from 250,000 to 300,000 tokens of prompt more than doubles the bill for GPT-6 Astra. Trimming a long prompt under 272,000 tokens saves more than moving to GPT-6.2 Astra. On price alone, GPT-6.2 Astra now undercuts its predecessor by 20% at every size.
Is GPT-6 Astra better than Claude Fable 5.1?
On most of OpenAI's own tests, yes, but not all of them. These rows are from the table at the end of OpenAI's launch post. OpenAI chose and ran the tests, reports "the maximum at any effort," and says ChatGPT results may differ "due to differences in the system prompts, tools available, etc." We haven't run them ourselves.
| OpenAI's test | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 |
|---|---|---|---|
| Terminal-Bench 4.0 (terminal tasks) | 57.9% | 37.3% | 55.8% |
| DeepSWE v1.1 (coding) | 74.1% | 72.7% | 67.4% |
| AutomationBench | 41.4% | 18.1% | 31.4% |
| BenchCAD (3D models from images) | 95.9% | 83.3% | 84.3% |
| GPQA Diamond (science) | 96.0% | 94.6% | 93.7% |
| FrontierMath Tier 4 (v2) | 97.6% | 83.0% | 87.8% |
| HealthBench Professional | 63.4% | 60.5% | 58.1% |
| Humanity's Last Exam, with tools | 57.2% | Not listed | 65.0% |
| Artificial Analysis Intelligence Index v4.1.1 | 61.2 | 60.9 | 65.7 |
The last two rows go to Anthropic's model, in OpenAI's own table. The post's opening rounds the FrontierMath score up to 98%; the table says 97.6%. On computer use, OpenAI puts Astra at 59.3% on Agents' Last Exam, against 55.5% for Claude Opus 5, and says it scored 72.6% on OSWorld 2.0 in about 40 minutes a task, where GPT-5.6 Sol took about 75 minutes for 65.7%.
How is GPT-6 Astra different from the other GPT-6 models?
It's the most expensive GPT-6 model on the API, and OpenAI deploys it with the misalignment monitoring it built for "Astra-class models." As of November 18, 2026, from OpenAI's model pages and changelog:
| Model | Released | Price per 1M tokens, in and out | Knowledge cutoff | "None" reasoning |
|---|---|---|---|---|
| GPT-6 Astra | September 3, 2026 | $10 and $50 | April 30, 2026 | No |
| GPT-6 Sol | September 22, 2026 | $2 and $10 | April 20, 2026 | Yes |
| GPT-6 Luna | September 22, 2026 | $0.10 and $0.50 | May 18, 2026 | Yes |
| GPT-6.1 Sol | September 29, 2026 | $2 and $10 | April 30, 2026 | No |
| GPT-6.2 Astra | November 17, 2026 | $8 and $40 | May 18, 2026 | No |
OpenAI's changelog describes GPT-6.1 Sol as "near-Astra performance" at "a lower cost than GPT-6 Astra," and its models page now sends new projects to GPT-6.2 Astra as "our flagship model." GPT-6 Astra keeps its own page and prices, and OpenAI's deprecations page gives it no shutdown date. For European developers, OpenAI's GPT-6 guide says Fast mode with EU data residency isn't available for GPT-6 Astra, while GPT-6.2 Astra supports US and EU residency on Standard, Fast, and Batch.
What does the GPT-6 Astra system card say?
That Astra is the first OpenAI model rated Critical in cybersecurity. The system card, published September 3, 2026, says that "with the right tools and access" Astra "can find previously unknown security flaws and develop new ways to exploit them" without a person guiding each step. It rates Astra High in biological and chemical capability and below High in AI self-improvement.
On data, the card is brief. Astra "was trained on diverse datasets and filtered through our data processing pipeline, including to reduce personal information." Some of its tests draw on real use. The hallucination section scores Astra on "de-identified ChatGPT conversations that users of our prior models have flagged as containing factual errors," and the safety benchmarks use conversations "representative of challenging examples from production data."
The card also reports three changes in how OpenAI watches the model:
- Harder to monitor. OpenAI found Astra "more capable of controlling its own CoT than GPT-5.6 Sol, and less likely to include incriminating information in its CoT," and able to evade its monitors in some tests that asked it to.
- Monitoring on every tool call. OpenAI added misalignment monitoring "to all tool-using inference involved in our external deployment of Astra."
- Stricter for some accounts. For users "flagged as potentially high risk," Astra's refusal boundary becomes "more conservative." OpenAI's safety update says Astra refused 91.5% of its cyber jailbreak tests, against 59% for GPT-5.6 Sol.
OpenAI's launch post adds that by default Astra "will refuse to comply with more advanced cybersecurity tasks such as creating proof-of-concept exploits," and that OpenAI plans to expand access with less restrictive safeguards through its Daybreak program for defensive work.
The card's change log records that on October 30, 2026, SecureBio corrected its biology score for Astra from 61.6% to 59.2%, because the first run had web search turned on and the comparison models didn't.
Can you use GPT-6 Astra on ChatGPT Plus?
Only in ChatGPT Work and Codex, not in regular Chat. The launch post said Astra would reach "all ChatGPT Plus, Pro, Business, and Enterprise users." OpenAI's help article on Astra usage, updated November 17, 2026, is narrower: "Plus includes Astra in Work and Codex, but not GPT-6 Pro in Chat." In Chat, "GPT-6 Pro is powered by Astra" on eligible Pro, Business, and Enterprise plans, which matches the Pro column of ChatGPT's pricing page: "Pro reasoning powered by GPT-6 Astra."
On Plus, OpenAI estimates 5 to 45 local Astra messages in each five-hour window of Work and Codex, against 350 to 3,000 for GPT-6 Luna, and weekly limits may apply as well. The same article says buying credits won't unlock Astra: "Credits pay for usage; they do not add model access." The GPT-6.2 rollout "does not add either model to regular Chat" either.
What happens to your GPT-6 Astra prompts in ChatGPT and the API?
The same model runs under two sets of rules, both tied to an account you hold with OpenAI. As of November 18, 2026:
| ChatGPT (Free, Plus, Pro) | OpenAI API | |
|---|---|---|
| Training on your prompts | On unless you turn off Improve the model for everyone (OpenAI) | "Not used to train or improve OpenAI models" unless you opt in (OpenAI) |
| Feedback buttons | Rating a reply can send "the entire conversation" to training, even after you opt out (OpenAI) | Not applicable |
| How long prompts are kept | Until you delete the chat, then deleted "within 30 days," with exceptions (OpenAI) | Abuse monitoring logs "for up to 30 days"; Responses API state 30 days by default (OpenAI) |
| A way to keep nothing | None on the pages we read; temporary chats aren't used for training while they stay temporary (OpenAI) | Zero Data Retention, "subject to prior approval by OpenAI" (OpenAI) |
| Astra's monitor | A paused task may ask you to review the action (launch post) | A stopped request returns HTTP 403, misalignment_policy_violation (OpenAI) |
In ChatGPT
Your Astra chats follow ChatGPT's normal data settings. OpenAI says one opt-out covers "ChatGPT conversations and Codex tasks," and the feedback rule above still applies after it (OpenAI). ChatGPT Business and Enterprise don't train on inputs or outputs by default, the same page says. We cover the full settings in does ChatGPT train on your data.
On the API
The launch post says Astra "supports Zero Data Retention for eligible API customers," and the data controls page adds that OpenAI reserves "the right to make models ineligible" for zero retention for specific customers, with written notice. Misalignment monitoring "reviews model reasoning and actions asynchronously" on Responses API requests; Chat Completions requests aren't covered by it. Using the API at all takes an organization, a key, billing, and code. Our page on whether the OpenAI API is private goes through the endpoints one by one.
How to use GPT-6 Astra without an OpenAI account
Every row in that table starts with an account at OpenAI: a ChatGPT login where training is on until you find the switch, or an API organization you fund and then ask OpenAI to approve for zero retention. On Morse, GPT-6 Astra comes with Pro, at $19 a month ($17 billed yearly), through its 250 monthly credits for the top models, alongside Claude Fable 5.1 and more. Max, at $49 a month, has 5,500 credits. You pick Astra from a menu, with no OpenAI account. Morse pricing explains the plans.
- Anonymous, to give GPT-6 Astra a hard problem without the provider knowing it's you. GPT models run in Anonymous mode. Our proxy removes your account, email, and IP address and sends the message under Morse's own account, so the provider sees the message without knowing it's you. It may keep it for a limited time to check for abuse, under its own policy, so leave out names and details you'd rather not send. Fable 5.1 wins two rows of OpenAI's own table and sits in Pro's credits too, so you can send the same prompt to each and judge the answers on your own work. Our guide to using ChatGPT anonymously explains the method.
- Private, for the follow-ups no host should keep. Open-weight models like DeepSeek, Kimi, and GLM, plus xAI's Grok, run in Private, the mode a new chat starts in. Morse has zero data retention agreements with its inference, search, and other providers, so they've agreed to keep nothing, and where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.
- E2EE, for drafts no company should read. Astra is closed-weight, so for these chats pick an open-weight model in a secure enclave. Your browser checks the enclave's hardware attestation, then encrypts the message so only that enclave can read it, and Morse's servers relay text they can't read. The guarantee rests on the enclave hardware and attestation working as designed. E2EE has no web search or image generation, since models can use only tools that run on the device. It comes with Pro.
Your chat history is stored in your browser, encrypted, not on Morse's servers. Morse doesn't train AI on your chats, sell your data, or show ads (privacy), and because it doesn't store chats, it usually has very little to give authorities. Clearing your browser or losing the device deletes your chats, and Morse can't recover them, so clear browser data only when you mean to.
Start free, then choose Pro at $19 a month to use GPT-6 Astra and Claude Fable 5.1 from one account.
Common questions
Is GPT-6 Astra the same as GPT-6 Pro in ChatGPT?
GPT-6 Pro in Chat is powered by Astra, OpenAI's help center says, but it has its own availability and message limits. Having Astra in Work and Codex doesn't include GPT-6 Pro in Chat, and Plus gets only the first.
Can you set temperature on GPT-6 Astra?
No. OpenAI's model page says GPT-6 Astra doesn't support custom temperature or top_p values or log probabilities, and it has no 'none' reasoning setting. The lowest effort is low and the highest is max.
Will GPT-6 Astra write exploit code?
Not by default. OpenAI's launch post says Astra refuses advanced cybersecurity tasks such as proof-of-concept exploits, and that less restrictive access comes through its Daybreak program for defensive work.
Is GPT-6 Astra being retired now that GPT-6.2 Astra is out?
Not as of November 18, 2026. OpenAI's models page points new projects to GPT-6.2 Astra, but gpt-6-astra keeps its own model page and prices, and the deprecations page lists no shutdown date for it.