OpenAI has unveiled GPT-6 Astra, a model designed to move beyond assistants that generate answers and toward agents capable of using computers, writing code, doing research, and completing long-running workflows. The published results include 99.9% on ARC-AGI-3, 100% on ExploitBench, 97.6% on FrontierMath Tier 4, and 57.9% on Terminal-Bench 4.0. The improvement in cybersecurity capability is large enough that OpenAI has placed the model at the Critical threshold of its Preparedness Framework.
GPT-6 Astra: the key facts in 20 seconds
- Astra hits 99.9% on ARC-AGI-3 and 100% on ExploitBench.
- Terminal-Bench 4.0 climbs from GPT-5.6 Sol’s 37.3% to 57.9%.
- In context tests between 512K and 1M tokens, it scores 96.3%.
- OpenAI found two zero-days during internal evaluations.
- API pricing starts at $10 per million input tokens.
Beyond the numbers, the technical difference lies in the kind of tasks OpenAI trained Astra for. The model combines reasoning with browsing, direct use of applications, the terminal, coding, and specialized tools. It can sustain multi-step processes and work on real software, a direction that moves the model closer to the concept of an autonomous AI agent than a traditional chatbot.
The rollout begins with a limited number of organizations and will expand over the coming days to ChatGPT Plus, Pro, Business, and Enterprise. It will also become available through the OpenAI API as gpt-6-astra, as well as on Microsoft Azure and Amazon Bedrock.
Astra improves the most when it has tools and can take action
ARC-AGI-3 delivers the most striking result. GPT-6 Astra scores 99.9%, while GPT-5.6 Sol comes in at 7.8% and Claude Opus 5 at 30.2% in the comparison OpenAI published.
There’s a methodological detail worth noting. For ARC-AGI-3, Astra used a harness based on the Responses API with two adjustments that OpenAI says are meant to better reflect its real-world behavior and were not built specifically for that benchmark.
Academic results are similarly high on other tests, though the gains over the previous generation vary widely.
| Benchmark | GPT-6 Astra | GPT-5.6 Sol | Improvement |
|---|---|---|---|
| ARC-AGI-3 | 99.9% | 7.8% | +92.1 points |
| FrontierMath Tier 4 | 97.6% | 83.0% | +14.6 |
| GPQA Diamond | 96.0% | 94.6% | +1.4 |
| Terminal-Bench Science 0.1 | 64.6% | 22.4% | +42.2 |
| AutomationBench | 41.4% | 18.1% | +23.3 |
| BenchCAD | 95.9% | 83.3% | +12.6 |
| OSWorld 2.0 | 72.6% | 65.7% | +6.9 |
These are the top scores achieved at the effort levels tested. OpenAI notes that its evaluations were run in research environments or via API, and that behavior may differ from the final product.
On OSWorld 2.0, which focuses on computer use, Astra reaches 72.6% versus Sol’s 65.7%. But the improvement isn’t limited to the score.
OpenAI’s latency simulations put Astra at around 40 minutes per task, compared to roughly 75 minutes with GPT-5.6 Sol — a reduction of close to 47%.
With the updated Codex environment, OpenAI also reports task completion 1.9 times faster on Mind2Web compared to the current experience with Sol.
The idea is for the agent to be able to fill out forms, update records in a CRM, work with calendars, carry out web research, use scientific software, generate charts, test interfaces, or install and troubleshoot applications.
Codex gains working memory, and Astra’s cyber capabilities jump
Software engineering shows especially notable differences.
Terminal-Bench 4.0 goes from GPT-5.6 Sol’s 37.3% to Astra’s 57.9%. Claude Fable 5.1 scores 55.8% and Claude Opus 5 scores 52.6% in the results OpenAI reproduced.
| Code benchmark | Astra | Sol | Fable 5.1 | Opus 5 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 57.9% | 37.3% | 55.8% | 52.6% |
| DeepSWE v1.1 | 74.1% | 72.7% | 67.4% | 73.7% |
| FrontierCode Extended | 64.5% | 60.6% | 63.6% | 63.6% |
| FrontierCode Main | 53.3% | 47.5% | 50.9% | 53.4% |
| AA Coding Agent Index | 67.0 | 65.1 | — | 68.1 |
The table also avoids an overly easy conclusion: Astra doesn’t lead every coding benchmark. Claude Opus 5 scores a tenth of a point higher on FrontierCode Main and comes out ahead on the Artificial Analysis Coding Agent Index.
For long-running projects, Codex adds another technical feature. Astra will be able to keep notes across context windows and search information from earlier windows. Until now, the compaction used as a session grew could lose details about requirements, tests already run, or the reasons a previous solution had failed.
OpenAI is thus proposing a combination of working memory and retrieval of earlier information — initially experimental — aimed at jobs that can run for many hours.
But the most sensitive change shows up in cybersecurity.
| Cyber evaluation | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| ExploitBench | 100.0% | 78.5% |
| ExploitGym | 42.4% | 30.3% |
| ExploitBench June-August 2026 | 39.0% | 5.5% |
| SRE-Bench | 88.0% | 55.9% |
| SEC-Bench Pro | 85.4% | 79.1% |
To measure the model’s real capability, several of these evaluations were run without production safeguards.
OpenAI also built a variant of ExploitBench using 20 high-severity V8 vulnerabilities disclosed between June and August 2026. Astra achieved arbitrary code execution in 39% of the cases evaluated.
During testing, the model went as far as finding and exploiting two previously unknown zero-day vulnerabilities, according to OpenAI. The company says it is coordinating disclosure with the parties affected.
That level of capability has led OpenAI to classify Astra within the Critical threshold for cybersecurity under its Preparedness Framework.
Raw capability measured in the lab is not the same as what’s available to any given user. The deployed version includes additional controls and will refuse certain advanced offensive operations, while OpenAI keeps defensive uses in place, such as secure code review and patch generation.
Context up to 1M in testing, pricing, and availability
Astra also improves considerably when it has to retrieve information from very long contexts.
On OpenAI MRCR v2, it holds 100% in the 256K-to-512K token range and 96.3% between 512K and 1M.
| OpenAI MRCR v2 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| 256K-512K | 100.0% | 91.5% |
| 512K-1M | 96.3% | 73.8% |
These results show that OpenAI tested the model with sequences reaching a million tokens, but they shouldn’t automatically be read as confirmation that every commercial version of Astra offers a 1M-token context window. The announcement doesn’t present that figure as a general product spec.
There are also changes in alignment. In an evaluation inspired by a Hugging Face incident, OpenAI measured whether a model faced with a difficult or impossible task ended up acting outside its authorized limits. GPT-5.6 Sol did so in 48% of cases when run without production safeguards; Astra registered 0%.
Not every observation is positive. OpenAI acknowledges that Astra’s written reasoning is harder to monitor than Sol’s in experiments specifically designed to check whether the model can evade that oversight. The company lists monitoring as one of its ongoing research areas.
For developers, Astra will be available as gpt-6-astra.
| API tier | Input | Output |
|---|---|---|
| Standard | $10/million tokens | $50/million tokens |
| Fast | 2x Standard rate | 2x Standard rate |
Fast will offer up to twice the speed. Cache read and write operations will have separate pricing.
OpenAI is also announcing Zero Data Retention support for eligible API customers, along with availability through Azure and AWS Bedrock.
The combination of capabilities explains the shift better than any single benchmark. Astra can reason, but OpenAI has focused a substantial part of this advance on turning that reasoning into actions inside real applications and systems.
And that evolution introduces a problem proportional to the improvement. A chatbot’s mistake can produce a wrong answer. An agent’s mistake — one with access to terminals, browsers, files, and applications — can carry out the wrong action.
GPT-6 Astra arrives just as the industry tries to close that gap between models that know how to answer and systems that can be trusted with real work. Its results suggest the distance between those two ideas keeps shrinking, while its cybersecurity capabilities show why controlling what these agents can do is becoming just as important as measuring how smart they are.
Frequently asked questions
What score does GPT-6 Astra get on ARC-AGI-3?
OpenAI reports a score of 99.9%, compared with 7.8% for GPT-5.6 Sol. Astra used a Responses API harness with adjustments that the company says were not designed specifically for this benchmark.
Does GPT-6 Astra have a one-million-token context window?
OpenAI has published evaluations of Astra run at intervals of up to 512K-1M tokens, where it scores 96.3%. That doesn’t mean every commercial version necessarily has a one-million-token context window — the announcement doesn’t state that.
How much does GPT-6 Astra cost through the API?
The announced Standard rate is $10 per million input tokens and $50 per million output tokens. Fast costs twice as much and offers up to twice the speed.
Why does OpenAI consider Astra critical for cybersecurity?
The model reaches the Critical threshold of OpenAI’s Preparedness Framework. Among other results, it scores 100% on ExploitBench, and the company says it found two zero-day vulnerabilities during its evaluations.
