OpenAI’s GPT-6 Astra, released on 3 September 2026, is its flagship reasoning model and the first above the GPT-5.6 Sol/Terra/Luna family. It delivers a genuine step change in computer use, agentic coding, and cybersecurity, but it is roughly flat on general intelligence and costs roughly 2.5x what GPT-5.6 Sol costs. This review breaks down the benchmarks, the pricing reality, and whether the upgrade is worth it for you.
What Is GPT-6 Astra?
GPT-6 Astra is OpenAI’s next-generation flagship large language model, succeeding GPT-5.6 Sol as the top end of the product line. OpenAI positions it as its “next generation of work” — the most intelligent and aligned model it has released, now available in ChatGPT Work, Codex, and via the API. It launched in limited preview on 3 September 2026, with paid users gaining access the following day and a rollout across ChatGPT Plus, Pro, Business, Enterprise, the OpenAI API, Microsoft Azure, and Amazon Bedrock.
Notably, Astra’s release was delayed after the Hugging Face incident in July 2026 so OpenAI could add more safety safeguards. It is a single dense reasoning model with no mini or nano tier — unlike the GPT-5.6 family, there is no lightweight sibling to serve as a cheap low-latency option.
GPT-6 Astra Specs and Pricing
The model is served under the API model ID gpt-6-astra. It accepts text and images as input and returns text. Below are the key specifications.
| Specification | GPT-6 Astra |
|---|---|
| Release date | 3 September 2026 |
| API model ID | gpt-6-astra |
| Context window | 1,050,000 tokens |
| Max input | 922,000 tokens |
| Max output | 128,000 tokens |
| Input price | $10 per million tokens |
| Output price | $50 per million tokens |
| Cached input | $1 per million tokens |
| Knowledge cutoff | 30 April 2026 |
| Reasoning effort levels | Low, medium, high, xhigh, max |
The headline 1,050,000-token context window is a total budget, not a free-for-all. Practical input is capped at 922,000 tokens, which matters if you plan to feed the model large codebases or long documents. There is also a 272K pricing cliff — past that point in the context budget you are billed differently — so teams doing long-context work should chunk against the lower number rather than the headline.
The knowledge cutoff is 30 April 2026, roughly a 10-week gain over the GPT-5.6 family’s 16 February 2026 cutoff. That matters if you work with very recent regulations, security advisories, or fast-moving technical documentation.
GPT-6 Astra Benchmarks Decoded
Benchmarks tell two different stories about Astra, and it is worth separating them.
Computer Use (OSWorld 2.0) — The Real Headline
The most important number in the whole release is on OSWorld 2.0, the computer-use benchmark. GPT-6 Astra scores 72.6% in about 40 minutes per task, versus GPT-5.6 Sol’s 65.7% in about 75 minutes. That is both a higher score and roughly 47% less time per task.
This is the true step change. It means Astra can complete UI tasks — filling forms, updating CRMs, managing calendars, drafting documents, running QA — on real applications, often without any API integration. OpenAI’s example is striking: completing Financial Modeling World Cup challenges about 4x faster than the human winner, purely via computer use. This is the “hand the agent a task” era versus the older “babysit the agent” era.
The ARC-AGI-3 Harness Controversy: 99.9% vs 62.7%
The most contested number is ARC-AGI-3, where OpenAI claims 99.9% for Astra. That number only holds under a stateful provider adapter — a harness that carries state and spends significant compute per task. Under an independent neutral harness, the score drops to 62.7%. Stateless API calls score far lower, in the 17–63% range depending on setup.
The honest read: the 99.9% figure is a marketing number that depends heavily on how the model is evaluated. Do not treat it as a general-intelligence score. Treat it as a measure of what the model can do with the right scaffolding and budget — which is real, but not what the headline implies.
Math, Science, and Agentic Coding
Astra saturates FrontierMath Tier 4 v2 at 97.6% and scores 100% on ExploitBench, OpenAI reports. On agentic coding, Terminal-Bench 4.0 shows Astra at 57.9% versus Sol’s 37.3% — a new high, at roughly 9% and 63% lower estimated cost per task than Sol and Claude Fable 5.1 respectively. These are the areas where the generational jump is real.
Where Astra Still Trails
Astra is not a clean sweep. On Humanity’s Last Exam (HLE) with tools, it scores 57.2% versus Claude Fable 5.1’s 65.0%. On the Artificial Analysis Intelligence Index, Astra scores about 61, essentially tied with Sol and below Claude Fable 5.1 and Opus 5. The honest summary from the research is: a real step change on agentic, computer-use, cyber, and math; roughly flat on general intelligence.
GPT-6 Astra vs GPT-5.6 Sol vs Claude Fable 5.1
| Metric | GPT-6 Astra | GPT-5.6 Sol | Claude Fable 5.1 |
|---|---|---|---|
| OSWorld 2.0 computer use | 72.6% (~40 min) | 65.7% (~75 min) | — |
| HLE with tools | 57.2% | — | 65.0% |
| Terminal-Bench 4.0 | 57.9% | 37.3% | Below Astra |
| Artificial Analysis Index | ~61 | ~61 (tied) | ~66 |
| Output price per M tokens | $50 | $20 | — |
| Context window | 1,050,000 | — | — |
The pattern is clear. Astra wins decisively on computer use, agentic coding, and math. Claude Fable 5.1 still leads on broad knowledge and general-intelligence-style tasks. And the price gap to Sol is significant — Astra’s output tokens cost 2.5x as much.
GPT-6 Astra in the Real World: Computer Use, Pro Work, and Cybersecurity (Daybreak)
Astra’s strongest enterprise signals come from the computer-use capability and early adopters. Reports cite Devin/Cognition using it for engineering work, Databricks using it on OfficeQA Pro for office-oriented question answering, and Hebbia noting that decks produced with Astra followed briefs about 17% more faithfully.
There is also a serious double-edged sword in cybersecurity. OpenAI reports Astra crosses the “Critical Preparedness” threshold — a level of offensive capability that raises real safety concerns. To manage this, exploit-creation capability is gated behind OpenAI’s Daybreak program rather than exposed freely. For enterprises, this cuts both ways: it means stronger defensive analysis, but it also means the tool carries real offensive risk and should be deployed with governance in mind.
Pricing and Cost-Per-Task Analysis — Is It Worth It?
At $10 input / $50 output per million tokens ($1 per million for cached input), Astra is roughly 2.5x the price of the model it replaces on output. That is the crux of the decision: you pay 2.5x for a narrow-band capability gain, so you need to be buying the right thing.
| Usage | Value of Astra |
|---|---|
| Computer-use automation on real apps | Highest — the OSWorld jump and time savings justify a premium |
| Agentic coding (Terminal-Bench) | High — better score AND lower cost per task |
| Long-document reasoning | Medium — real deep-context retention, but watch the 272K pricing cliff |
| General Q&A / drafting | Low — flat gain over Sol, and Claude Fable 5.1 is better for less |
The practical takeaway: if your work is computer-use automation or agentic coding, the 2.5x premium buys a real, measurable improvement. If you mostly use an LLM for general reasoning and drafting, the premium is wasted — you are paying more for roughly the same capability you already had.
How to Access GPT-6 Astra
GPT-6 Astra is broadly available. You can use it through:
- The OpenAI API under the model ID
gpt-6-astra - ChatGPT (Plus, Pro, Business, and Enterprise plans)
- ChatGPT Work and Codex
- OpenAI’s GPT-6 Astra Pro tier
- Amazon Bedrock and Microsoft Azure
For developers, the main migration gotcha is the context budget: the 1,050,000-token window is a total cap with input effectively limited to 922,000 tokens, and the 272K pricing cliff sits below the headline — chunk your long-context workloads accordingly.
Verdict: Who Should (and Shouldn’t) Switch
Switch to GPT-6 Astra if you build computer-use automation, run heavy agentic coding, or need class-leading math and deep-context retention — the capability gain over Sol is real and the cost-per-task on agentic work is actually lower.
Hold back if your workload is general drafting, Q&A, or broad reasoning. You will pay 2.5x Sol’s price for roughly flat performance, and Claude Fable 5.1 remains a better value on the intelligence axis. And treat every ARC-AGI-3 claim with the harness caveat firmly in mind: 99.9% is real only with the right scaffolding; the neutral-harness number is 62.7%.
FAQ
What is GPT-6 Astra?
GPT-6 Astra is OpenAI’s flagship reasoning model released 3 September 2026, succeeding GPT-5.6 Sol. It is a single dense model with a 1,050,000-token context window and a step change in computer-use and agentic-coding ability.
How much does GPT-6 Astra cost?
Astra costs $10 per million input tokens and $50 per million output tokens, with cached input at $1 per million. That is roughly 2.5x the output price of GPT-5.6 Sol.
Is GPT-6 Astra better than Claude Fable 5.1?
It depends on the task. Astra wins on computer use (OSWorld 72.6%) and agentic coding, but Claude Fable 5.1 still leads on general intelligence tasks like Humanity’s Last Exam (65.0% vs 57.2%).
Is the GPT-6 Astra ARC-AGI-3 score of 99.9% real?
Only under OpenAI’s stateful provider-adapter harness with significant compute. Under an independent neutral harness, Astra scores 62.7% on ARC-AGI-3, and stateless API calls score 17–63%.
What is the GPT-6 Astra context window?
The context window is 1,050,000 tokens total, with input capped at 922,000 tokens and output at 128,000 tokens. Note there is a 272K pricing cliff below the headline context.
When was GPT-6 Astra released?
GPT-6 Astra launched in limited preview on 3 September 2026, with paid users gaining access the following day and a broader rollout across ChatGPT, the API, Azure, and Amazon Bedrock.