GPT-6 Astra: What Really Changed

Every model launch comes with a chart showing the new model on top. Astra’s chart has a shape to it. It leapt on some tests and barely moved on others, and the split runs along a clear line.

OpenAI calls GPT-6 Astra the most intelligent and aligned model in the world. Its benchmark tables tell a narrower story. Astra got dramatically better at working through a job step by step, whether that is operating software, learning an unfamiliar task, proving a hard theorem or finding a security hole. On the tests that are closer to recall, it gained a point or two.

What is GPT-6 Astra, in plain terms?

It is the model rolling out behind ChatGPT for paying users over the days after launch, the successor to GPT-5.6 Sol. If you have used ChatGPT this year, Astra is the next version of the thing you already talk to.

The word OpenAI keeps using about it is agent. An AI agent takes a goal, works through the steps, uses software, and reports back. We wrote about this shift when models first started coordinating with each other, and Astra is the clearest example of it so far.

Two behaviour changes matter more than any score. Astra asks you a focused question when the answer could change the outcome, instead of guessing and writing three pages of the wrong thing. And when you redirect it mid-task, it folds the new instruction into the original job rather than treating your correction as a brand new assignment.

The leap is in doing, not knowing

On ARC-AGI-3, a test of how quickly a model learns an unfamiliar interactive task, Astra scores 99.9%. GPT-5.6 Sol scores 7.8%. The average human tester scores 48%. Greg Kamradt of the ARC Prize Foundation, which runs the benchmark, called the result a meaningful step change in frontier-model performance. A footnote records that Astra ran on a harness with two settings changed, which OpenAI says better reflects real-world performance.

The recall-style tests tell a different story. On GPQA Diamond, a set of graduate-level science questions, Astra scores 96.0% against Sol’s 94.6%, a rounding error next to a 92-point swing. ARC-AGI-1 moves one point. The pattern repeats across the tables: large gains wherever the model has to work through something, small gains wherever it only has to know the answer.

Maths is the loud exception, and it belongs on the working-through side of the line. FrontierMath Tier 4, a set of research-level problems, goes from Sol’s 83.0% to Astra’s 97.6%. Those problems are not recalled, they are worked out over many steps, which is the thing that improved.

On the independent Artificial Analysis Intelligence Index v4.2, published the day after launch, Astra comes second behind Claude Fable 5.1 while costing roughly half as much per task as the three Claude models it is measured against. Worth knowing which direction that comparison runs: Astra is cheaper per task than Claude, not cheaper than OpenAI’s own GPT-5.6 Sol.

What can it do on a computer?

OpenAI’s demonstrations are ordinary office work rather than research problems:

• Reading a W-2 and filling in a Form 1040 in the browser, leaving a draft return for a person to check
• Building and refining a dashboard inside Power BI, turning vehicle data into comparisons of range, price and efficiency
• Formatting a legal memo and a will in a word processor, handling the headings and page layout
• Running a website’s quality-assurance checklist, clicking through navigation and search to find what is broken
• Laying out a printed circuit board in KiCad, and modelling a five-speed transmission in FreeCAD

On OSWorld 2.0, a computer-use test, Astra scores 72.6% at roughly 40 minutes per task. Sol scores 65.7% at roughly 75 minutes.

OpenAI’s own table has Claude ahead on two measures

Two numbers in OpenAI’s published comparison table go the other way.

Publishing a number your rival wins on is a point in OpenAI’s favour, and it tells you something useful. Astra is not a model that knows more than everything else. It is a model that works a computer better than anything else, more cheaply, and with fewer wrong turns. Which one you want depends on the job. Our Claude versus ChatGPT comparison goes deeper on that split, and the 2026 model overview tracks where each family currently stands.

What does it cost, and who gets it?

If you subscribe to ChatGPT, Astra is included in the allowance you already pay for, and you can buy credits for more usage on top. Astra is on every paid tier. Pro, Business and Enterprise also get a higher-capability version called GPT-6 Astra Pro.

One detail worth passing to your IT department: on Enterprise plans, access is off by default at launch. An administrator has to switch it on for the workspace. If your company says it has not got Astra yet, that is probably why.

Developers get it as gpt-6-astra in the OpenAI API, and through Microsoft Azure and AWS Bedrock. Pricing is $10 per million input tokens and $50 per million output tokens, where a token is roughly three-quarters of a word. A Fast mode runs up to twice as quickly for twice the price.

Astra hit OpenAI’s own Critical risk threshold

OpenAI states plainly that Astra meets the Critical threshold for cybersecurity under its Preparedness Framework, the internal scale it uses to rate how dangerous a model’s capabilities are. On ExploitBench, which tests turning known software flaws into working attacks, Astra completed every task with its production safeguards switched off. Sol managed 78.5% under the same conditions. During that testing Astra found and used two previously unknown vulnerabilities, which OpenAI says it is disclosing to the maintainers.

The version being released refuses the more advanced offensive requests, such as building working demonstrations of how to break into a system. That is the position for now: OpenAI says it plans to loosen those safeguards within weeks for approved defensive users. It also warns that the extra safety checks can pause or stop legitimate work. In ChatGPT you may be asked to review an action before it continues, and in the API the task stops.

OpenAI also found that Astra’s written reasoning is harder to monitor than Sol’s when tested with prompts that asked it to evade monitoring. That reasoning is the running set of notes a model writes to itself while working, which safety researchers read to check what it is up to. OpenAI puts this down to Astra having more control over its written reasoning on simple tasks and solving problems in fewer steps, and adds that it still appears to struggle to hide the reasoning that hard tasks require. It calls the drop serious and lists it as an open research problem.

What this does not change for you

• It fills in the tax form. Whether the numbers are right, and whether the return is worth filing, is still your call and your signature.
• The legal memo is formatted in a minute. What position it takes is still a decision only you can put your name to.
• It builds the dashboard. Knowing which comparison your board needs is judgement, and judgement is the scarce part.
• A model that scores 99.9% on learning unfamiliar tasks still has nothing riding on whether your work turns out well.

The tedious middle of the work just got cheaper. That pushes the value of your day toward the parts that need a person.

Questions readers are asking

Do I have to do anything to get it?

No, if you are on Plus, Pro or Business. It becomes available over the days following launch. On Enterprise, an administrator has to turn it on first.

Is it better than Claude?

At operating software, yes, though the size of the win depends on who you compare it to. Astra beats GPT-5.6 Sol by huge margins. Against Claude Opus 5 on the two computer-use tests where OpenAI publishes a Claude score, the gap is a couple of points. On raw reasoning it loses: all three Claude models in OpenAI’s table beat Astra on the Artificial Analysis v4.1.1 composite and on Humanity’s Last Exam. The useful question is which task you need done.

Should I let it use my computer?

Start with work you can check. Data entry, formatting, filling forms from documents you already have. Keep it away from anything involving money moving, credentials or irreversible sends until you have watched it work a few times.

What is GPT-6 Astra Pro?

OpenAI has not said what separates it, beyond the name and who gets it: Pro, Business and Enterprise plans. Plus subscribers get standard Astra.

Did it really solve open maths problems?

Yes, and we covered that story on its own when it happened. See what Astra solved, including why the machine-checkable proofs matter more than the answers.

Why does the cost matter if I pay a flat subscription?

Because cheaper tasks mean your allowance stretches further before you hit a limit or buy credits. On the API it matters directly, since you pay per token.

References

• OpenAI’s launch announcement, GPT-6 Astra: A new generation of intelligence, which contains every benchmark table quoted above
• Artificial Analysis Intelligence Index v4.2, the independent composite score and the cost-per-task chart
• ARC Prize Foundation, which runs the ARC-AGI benchmarks

Related guides

• ChatGPT dots: always-on agents that run on Astra
• OpenAI Astra solved 10 maths problems
• Claude versus ChatGPT in 2026
• The AI models of 2026
• When your AI assistants start talking to each other
• The everyday AI agent team
• What you get on Anthropic’s free tier
• The AI glossary