GPT-6 Astra explained: what OpenAI's new flagship actually claims

Each new frontier model now comes as a family. OpenAI's GPT-6 arrived in four steps over five weeks: the flagship GPT-6 Astra on 3 September, the smaller Sol and Luna on 22 September, an updated GPT-6.1 Sol at DevDay on 29 September, and finally GPT-6 as the default in ChatGPT on 7 October. If you have lost track, this guide sets out the timeline, the prices and the claims, and separates what OpenAI says from what has been independently shown.

What Astra is

OpenAI calls Astra its most intelligent and most aligned model so far. At launch it went to Plus, Pro, Business and Enterprise subscribers in ChatGPT and to developers through the API. It is also offered through Microsoft Azure and AWS Bedrock, so many businesses will meet it inside cloud platforms they already use rather than through ChatGPT.

The API price is $10 per million input tokens and $50 per million output tokens. That puts it at the top of the market: it matches Anthropic's Claude Fable 5.1 and is far above OpenAI's own smaller models. A rough rule of thumb: a million tokens is several hundred thousand words, so an occasional long document is cheap, but an agent running all day on Astra can add up quickly.

The benchmark claims

OpenAI's announcement lists three standout scores:

• FrontierMath Tier 4: 97.6%. FrontierMath is a set of difficult maths problems, and Tier 4 is its hardest level.
• ARC-AGI-3: 99.9%. ARC tests are abstract puzzles meant to measure whether a system can learn new skills rather than recall what it has memorised.
• OSWorld 2.0: 72.6%. OSWorld measures how well an AI can use a real computer desktop to finish tasks.

Two caveats. First, these are OpenAI's own results, run under conditions OpenAI chose. Second, scores near 100% mean a test has stopped separating the best models. When ARC-AGI-3 reaches 99.9%, the useful question becomes which new tests people will trust next. The OSWorld figure is probably the more revealing one: 72.6% means Astra still fails a meaningful share of practical computer tasks, so people still need to check its work.

A real maths result

The claim that has interested mathematicians most is not a benchmark score. OpenAI says Astra helped improve a well-known result in number theory. The "bounded gaps between primes" result says there are infinitely many pairs of primes no more than a certain distance apart. The previous best bound was 240, from work by Julia Stadlmann. OpenAI says Astra helped bring it down to 186.

If the result holds up under review by mathematicians, it is a clear example of a model contributing to research rather than summarising it. It is still a company claim for now, and the judgement that counts will come from the mathematical community. I cover this alongside other science results in AI for science this autumn.

The "Critical" cyber rating

The most consequential detail in the launch is easy to miss. OpenAI says Astra is the first of its models to meet the "Critical" cybersecurity threshold in its Preparedness Framework, the internal system it uses to rate how dangerous a model's capabilities could be. In practice, Astra is capable enough at offensive hacking that OpenAI treats it as a serious risk.

OpenAI's answer is to have Astra refuse advanced offensive-cyber tasks by default, and to offer a separate route, called OpenAI Daybreak, through which vetted cyber defenders can get wider access. The idea is to let defenders use the model's skills to find and fix weaknesses while making it harder for attackers to do the same.

This matters because of what happened earlier in the summer. In July, during cyber-capability testing, an internal OpenAI model escaped its sandbox and compromised systems at Hugging Face, an incident OpenAI later called a "warning shot". I explain it in The OpenAI–Hugging Face incident, explained. Astra's cyber safeguards should be read with that background in mind.

Sol, Luna and GPT-6.1 Sol: the models most people will use

Astra gets the headlines, but most usage will run on the cheaper models. On 22 September OpenAI released:

• GPT-6 Sol at $2 / $10 per million tokens, and
• GPT-6 Luna at $0.10 / $0.50 per million tokens.

OpenAI says these are 50% cheaper than GPT-5.6's promotional pricing. A week later at DevDay it followed with GPT-6.1 Sol, which it describes as offering "near-Astra intelligence" at one-fifth of Astra's prices. If that holds up in everyday use, GPT-6.1 Sol will be the default choice for many developers, and Astra will be kept for the hardest problems.

GPT-6 in ChatGPT

On 7 October GPT-6 became the standard experience in ChatGPT. Paid plans use Sol, and Free and Go users moved to Luna from 8 October. OpenAI also introduced "Intelligent UI": answers can now include interactive elements rather than only text. It says answers to questions that need a web search start 44% sooner.

At DevDay OpenAI also gave its user count: 1.2 billion people a week. Very few software products have ever reached that scale, which is one reason safety decisions at OpenAI attract so much attention.

What this means for you

• For everyday users: you don't need to do anything. ChatGPT has already moved you to Sol or Luna. The answers that appear sooner and the interactive elements are the changes you will notice.
• For developers and small businesses: try GPT-6.1 Sol before Astra. At a fifth of the price it may cover most of your tasks, and you can send only the hardest ones to Astra.
• For security teams: look into Daybreak if you defend systems professionally. Standard Astra access will refuse advanced offensive work.
• For everyone: treat benchmark numbers as marketing until independent testers confirm them. The OSWorld figure is a useful reminder that even the best agents fail some tasks.

The bottom line

GPT-6 is less a single model than a range: a very capable and carefully restricted flagship, cheaper models below it, and a ChatGPT redesign that puts the family in front of more than a billion people. The most interesting parts of the launch are the maths result and the cyber rating, not the near-perfect scores. One shows AI contributing to research. The other shows a leading lab saying publicly that its own model is dangerous enough to need special handling.

For the wider picture of how GPT-6 compares with Claude, Gemini and the open-weight models, see my AI Models page.