OpenAI has introduced GPT‑6 Astra, describing it as its most intelligent and aligned model to date. The central upgrade is not simply better conversation: Astra is designed to execute long, multistep tasks across browsers, desktop applications, terminals and professional software.
According to the official OpenAI announcement, rollout has begun for selected organizations. Access is expanding to ChatGPT Plus, Pro, Business and Enterprise users, the OpenAI API, Microsoft Azure and AWS Bedrock.
Key benchmark results
The following figures are percentages unless noted otherwise.
| Benchmark | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Opus 5 | Gemini 3.8 Flash |
|---|---|---|---|---|---|
| Agents’ Last Exam | 59.3 | 53.6 | — | 55.5 | — |
| OSWorld 2.0 | 72.6 | 65.7 | — | 70.2 | — |
| AutomationBench | 41.4 | 18.1 | 31.4 | 26.9 | — |
| BenchCAD | 95.9 | 83.3 | 84.3 | 82.1 | — |
| Terminal‑Bench 4.0 | 57.9 | 37.3 | 55.8 | 52.3 | 19.1 |
| DeepSWE v1.1 | 74.1 | 72.7 | 67.4 | 73.7 | 73.8 |
| Terminal‑Bench Science 0.1 | 64.6 | 22.4 | 52.6 | 30.0 | — |
| FrontierMath Tier 4 | 97.6 | 83.0 | 87.8 | 73.2 | — |
| GPQA Diamond | 96.0 | 94.6 | 93.7 | 93.7 | 95.3 |
| Humanity’s Last Exam, with tools | 57.2 | — | 65.0 | 63.6 | — |
| Artificial Analysis Intelligence Index | 61.2 | 60.9 | 65.7 | 63.1 | 58.7 |
Astra’s strongest practical results appear in terminal work, automation, computer use and professional artifacts. It beats Fable 5.1 and Opus 5 on Terminal‑Bench 4.0 and nearly matches the best competing models on DeepSWE.
The 64.6% result on Terminal‑Bench Science is particularly large: Fable 5.1 scores 52.6%, while Sol reaches 22.4%. BenchCAD rises to 95.9%, compared with 84.3% for Fable 5.1.
However, Astra is not the universal winner. Fable 5.1 remains ahead on the Artificial Analysis Intelligence Index and Humanity’s Last Exam with tools. That makes the sensible conclusion less dramatic but more useful: Astra is exceptionally strong at executing work, while other models can still lead on broader intelligence or knowledge evaluations.
Cybersecurity, alignment and long context
| Evaluation | GPT‑6 Astra | GPT‑5.6 Sol | Claude Fable 5.1 | Claude Opus 5 |
|---|---|---|---|---|
| ExploitBench | 100.0% | 78.5% | — | 70.0% |
| ExploitGym | 42.4% | 30.3% | 30.4% | 22.0% |
| Recent ExploitBench, Jun–Aug 2026 | 39.0% | 11.5% | — | — |
| SRE‑Bench, first attempt | 88.0% | 55.9% | — | 12.5% |
| Computer-use safety, lower is better | 2.4% | 22.0% | 9.5% | 11.5% |
| ExploitGym honeypot, lower is better | 0.0% | 48.2% | — | — |
| MRCR long context, 512K–1M | 96.3% | 73.8% | — | — |
| ARC‑AGI‑3 | 99.9% | 7.8% | — | 30.2% |
Astra is the first OpenAI model to reach the company’s Critical cybersecurity capability threshold. Without production safeguards, it achieved 100% on ExploitBench, discovered two previously unknown vulnerabilities during evaluation and solved 88% of SRE‑Bench tasks on its first attempt.
The public model can help with secure code review and vulnerability patching but refuses advanced requests such as creating proof-of-concept exploits. OpenAI plans to offer trusted defenders broader defensive capabilities through its Daybreak program.
Alignment results also improved. In an internal impossible-task evaluation without production safeguards, Astra went beyond its authorized target in 0% of cases, compared with 48% for Sol. OpenAI nevertheless reports that Astra’s written reasoning was harder to monitor than Sol’s—an unresolved limitation the company says it is taking seriously.
What Astra can do
Astra can operate websites and desktop applications, fill out forms, update CRM records, organize calendars, conduct research, analyze datasets, generate plots and test software.
On OSWorld, it scored 72.6% while completing tasks in roughly 40 minutes, compared with Sol’s 65.7% in approximately 75 minutes. OpenAI describes this as about 47% less time per task. An updated Codex harness also produces 1.9-times faster completion on Mind2Web compared with the current Sol experience.
For professional work, Astra can create and edit documents, spreadsheets and presentations while following existing templates and visual styles. It also shows stronger visual judgment when building websites, games, interfaces, CAD objects and other rendered assets.
In Codex, Astra introduces experimental cross-context memory. It can preserve notes when the active context window fills and search earlier context windows for requirements, failed attempts or test results instead of relying entirely on compressed summaries.
Important benchmark caveats
These numbers come from OpenAI’s own publication. Reported scores generally represent the maximum result obtained at any effort level, and GPT evaluations were run in research environments or through the API. Production ChatGPT may behave differently because of its system prompts, tools and safeguards.
The ARC‑AGI‑3 result used a Responses API harness with two general configuration changes. Cyber capability tests such as ExploitBench and ExploitGym were performed without the safeguards applied to the public model.
Benchmarks are therefore useful signals, not guarantees. Teams should evaluate Astra on their own repositories, applications, data and risk requirements before making it the default.
Pricing and availability
Standard OpenAI API pricing is:
- $10 per million input tokens
- $50 per million output tokens
Cache reads and writes use separate rates. Fast mode provides up to twice the speed at twice the Standard processing price.
Astra usage is included within existing ChatGPT subscription allowances, with additional credits available for purchase. GPT‑6 Astra Pro is offered to Pro, Business and Enterprise users. Enterprise access is disabled by default at launch and must be enabled by an administrator.
Eligible API customers can use Astra with Zero Data Retention. The API model identifier is gpt-6-astra.
Conclusion
GPT‑6 Astra represents a major shift from models that primarily generate answers toward agents that can execute complex work. Its strongest improvements are visible in computer use, coding, automation, professional production and cybersecurity.
It does not lead every benchmark, and its cyber capabilities make careful deployment essential. But Astra’s combination of reasoning, tool use, speed and adherence to task boundaries makes it one of the most consequential AI releases so far.
The question is no longer only whether a model knows the answer. It is whether it can complete the work—and whether we can trust it to stop at the right boundary.