GPT-6 Astra is OpenAI's frontier model, successor to GPT-5.6 Sol, built for hard reasoning, coding and, above all, autonomous computer use. A 1,050,000-token context window, a single Pro sibling, and a September 3, 2026 rollout across paid ChatGPT plans, the API, Microsoft Azure and AWS Bedrock. On the API side, input starts around $10 per million tokens. One model, one job description, and that job is everything you do on a screen.
GPT-6 Astra sets a new state of the art for computer use, browsing, software engineering and professional work. In practice, the model takes over a screen, fills out web forms, updates CRM records, installs software or runs QA checks on a website before it goes live. On OSWorld 2.0, the reference benchmark for this kind of work, it scores 72.6% while cutting time per task by roughly 47% compared with its predecessor.
You describe the mission, it opens the browser, types, navigates and comes back with the finished job. Five reasoning effort levels, from low to max, let you dial thinking depth up or down (and your bill along with it). The refreshed Codex tooling speeds up coding sessions too, with OpenAI citing 1.9x faster task completion on Mind2Web.
GPT-6 Astra is the first OpenAI model rated Critical for cybersecurity under the company's Preparedness Framework. Put plainly, it can uncover previously unknown flaws, including zero-day vulnerabilities, which is why reinforced safeguards can pause work the system finds suspicious.
The published scores read like the kind of sheet tracked by AI model leaderboards. Astra saturates FrontierMath Tier 4 at 97.6%, sits near perfection on ARC-AGI-3, and ten previously open math problems fell before launch, with proofs formalized in Lean.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| DeepSWE v1.1 (agentic coding) | 74.1% | 70.8% |
| Cyber jailbreak refusals | 91.5% | 59% |
| Boundary task solved in one attempt | 88.0% | 55.9% |
| Boundary task solved within four attempts | 99.2% | 68.7% |
GPT-6 Astra is available across ChatGPT Plus, Pro, Business and Enterprise plans, through the API under the gpt-6-astra ID, and via Microsoft Azure and AWS Bedrock. An Astra Pro variant ships with the Pro, Business and Enterprise plans. No Luna, Terra or Sol siblings this time; the lineup stops at two models.
API rates sit at $10 per million input tokens and $50 per million output tokens, with cached input at $1, Batch and Flex processing at half price, and a Fast mode at double the rate for roughly 2.5x the speed. Past 272,000 tokens of context, multipliers kick in. Those figures date from launch day, so treat OpenAI's pricing page as the only document that ages well.
No. Access requires a ChatGPT Plus, Pro, Business or Enterprise subscription, and the API bills per token. Free ChatGPT accounts were left out of the launch rollout. Some third-party platforms expose the model for trial runs, but official access goes through a paid OpenAI account or a partner cloud.
Both models carry the same list price on the API. Anthropic's Fable 5.1 earned its reputation on long-haul coding sessions, while Astra pulls ahead the moment a screen, a browser or desktop apps have to be driven. For automating complete computer-based tasks, Astra is the favorite.
Safety monitoring runs alongside the model because of its Critical cybersecurity rating. In ChatGPT and Codex you get a review prompt, while in the API a flagged task can halt outright. OpenAI admits the filters still fire on harmless work and says it keeps tuning them.
Zero Data Retention is supported by Astra on the API side, a meaningful point for companies bound by confidentiality rules. OpenAI's models also go through a training-data filtering pipeline aimed at reducing personal information.
Verdict: The bill stings, the delegation delivers. Developers, analysts and anyone glued to a browser now have a model able to finish the job on its own, from web form to final report.
Claude Fable 5.1 
Claude Fable 5 
Grok 4.6 
