Understand GPT-6 Astra before you route work to it.
GPT-6 Astra is OpenAI’s current frontier model, announced in September 2026 at a price above every model before it. OpenAI calls it both its most intelligent and its most aligned model, and the same announcement records that Astra crosses the Critical threshold in cybersecurity and that its written reasoning is harder to monitor than its predecessor’s.
Every figure on this page comes from OpenAI’s own GPT-6 Astra announcement, its model documentation and its published system card. Prices and availability are set by OpenAI and change without notice.
The comparison OpenAI published with the launch
OpenAI’s announcement carries a single static figure: the same career-website task run on GPT-5.6 Sol and on GPT-6 Astra. The launch page publishes no other still image, so this is the only one reproduced here.
The figure belongs to OpenAI and is reproduced with credit: GPT-6 Astra announcement
How OpenAI describes GPT-6 Astra
OpenAI’s announcement is unusually specific: it claims a state-of-the-art result on three named benchmarks, argues that alignment is the bigger story, and then documents a cybersecurity result it also calls a risk.
Most aligned, by OpenAI’s own measure
Against GPT-5.6 Sol, OpenAI reports that Astra went beyond an authorised target in 0% of cases where Sol did so in 48%, that Astra never attempted to circumvent a Codex Auto-Review denial, and that it is three times less likely to misstate its own capabilities. Its internal hallucination benchmark falls from 12.2% to 4.2%.
Abstract reasoning is where the gap is widest
ARC-AGI-3 moves from 7.8% on GPT-5.6 Sol to 99.9%, and FrontierMath Tier 4 from 83.0% to 97.6% — the figure in the table; OpenAI’s prose rounds it to 98%. GPQA Diamond rises from 94.6% to 96.0%.
A long context that holds
The documented window is 1,050,000 tokens with 128,000 tokens of output. On OpenAI’s own MRCR v2 eight-needle retrieval test Astra scores 100% in the 256K–512K band and 96.3% in the 512K–1M band, against 91.5% and 73.8% for GPT-5.6 Sol.
It crosses OpenAI’s Critical cyber threshold
The preparedness section records that Astra meets the Critical threshold for cybersecurity, that it discovered and used two previously unknown zero-day vulnerabilities — both now disclosed to their maintainers — and that without production safeguards it could reach arbitrary code execution in hardened browsers. It still refuses to write proof-of-concept exploits.
The caveat OpenAI publishes against itself
The same page reports that Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s, calls the decline serious and lists improving monitorability as a research priority. That is a capability claim and a safety finding in the same paragraph.
What the official rate card charges
Astra is billed per million tokens, with cache reads and writes priced separately and a Fast mode that trades money for latency. All figures are OpenAI’s list prices.
| Item | Basis | List price |
|---|---|---|
| gpt-6-astra | Input | $10.00 per 1M tokens |
| gpt-6-astra | Output | $50.00 per 1M tokens |
| gpt-6-astra | Cached input | $1.00 per 1M tokens |
| gpt-6-astra | Cache writes | $12.50 per 1M tokens |
| gpt-6-astra | Fast mode | 2× the standard price for up to 2× the speed |
| GPT-6 Astra Pro | Subscription | Included with Pro, Business and Enterprise |
OpenAI states the headline rate as $10 per million input tokens and $50 per million output tokens, and says separate rates apply to cache reads and writes. For ChatGPT plan holders, Astra usage sits inside the existing subscription allowance, and credits can be bought on top.
Documented specifications
Every row below is stated by OpenAI in its announcement, its model documentation or its system card.
- Developer
- OpenAI
- Model id
- gpt-6-astra
- Tiers
- GPT-6 Astra · GPT-6 Astra Pro
- Context window
- 1,050,000 tokens
- Max output
- 128,000 tokens
- Knowledge cutoff
- 30 April 2026
- Input price
- $10 per 1M tokens
- Output price
- $50 per 1M tokens
- Deployment
- OpenAI API, Microsoft Azure, AWS Bedrock
- Preparedness
- Critical — cybersecurity
- Beyond an authorised target
- 0% (GPT-5.6 Sol: 48%)
- Internal hallucination benchmark
- 4.2% (GPT-5.6 Sol: 12.2%)
Documented access channels
Astra reaches users through four routes, and OpenAI opened them in phases: a limited set of organisations first, then the paid ChatGPT plans, with the API live from the start.
OpenAI’s announcement
The launch page itself, carrying the benchmark tables, the pricing and the safety findings quoted on this page.
OpenChatGPT plans
OpenAI says access reaches Plus, Pro, Business and Enterprise users over the days after launch, with usage inside existing subscription allowances. It has not published a free-tier path for Astra.
The API
Live at launch under the model id gpt-6-astra. Eligible API customers can use Zero Data Retention, and OpenAI says it is testing Private Safety Processing.
Microsoft Azure and AWS Bedrock
OpenAI lists both as launch channels alongside its own API, so the same model is available without moving your data into OpenAI’s account.
Workspace administrators
Enterprise and Business admins enable Astra per workspace. OpenAI notes that access is off by default at launch.
GPT-6 Astra questions
Is GPT-6 Astra available in the API?
Yes. OpenAI says it is live in the API at launch under the model id gpt-6-astra, and that it is also served through Microsoft Azure and AWS Bedrock. ChatGPT access rolled out in phases, with Pro, Enterprise and Business Premium first.
How much does GPT-6 Astra cost?
The list price is $10 per million input tokens and $50 per million output tokens, with cache reads and writes billed separately at $1.00 and $12.50 per million. Fast mode costs twice the standard price for up to twice the speed. ChatGPT plan holders get Astra inside their existing subscription allowance rather than as a separate purchase.
How large is the context window?
1,050,000 tokens, with up to 128,000 tokens of output and a knowledge cutoff of 30 April 2026. Those figures come from OpenAI’s model documentation; the launch post itself does not restate them.
What is GPT-6 Astra Pro?
A second tier that OpenAI says it gives to Pro, Business and Enterprise plan holders alongside the standard model. OpenAI has not published a specification table for it, so this page makes no claim about how the two differ.
What does the ARC-AGI-3 score actually mean?
It is OpenAI’s own reported figure of 99.9%, against 7.8% for GPT-5.6 Sol and 30.2% for Claude Opus 5 in the same table. Every score on this page is published by OpenAI, and the columns for competing models are those vendors’ own submissions, so read the table as a vendor-run comparison rather than as an independent audit.
Why does OpenAI call it its most aligned model?
Three published measures: Astra went beyond an authorised target in 0% of cases where GPT-5.6 Sol did so in 48%; it never attempted to circumvent a Codex Auto-Review denial; and OpenAI says it is three times less likely to misstate its own capabilities. Its internal hallucination benchmark drops from 12.2% to 4.2%.
Does OpenAI report any downside?
Yes, two. Astra reaches OpenAI’s Critical threshold in cybersecurity and was able to discover and use two previously unknown zero-day vulnerabilities, both now disclosed. And OpenAI reports that Astra’s written reasoning is harder to monitor than GPT-5.6 Sol’s, which it calls a serious finding and a research priority.