Mathematicians spend careers chasing a single open problem. Some questions in group theory and combinatorics have sat unanswered for so long that entire subfields grew up around trying, and failing, to close them. On August 1, OpenAI said an internal, unreleased model had closed ten of them in one batch, and that the whole exercise cost roughly what a decent laptop does.
The model is called Astra. It has not shipped as a product, has no public price, and no confirmed release date. What it does have is a 249 page manuscript, a set of machine checked proofs sitting on GitHub, and a growing argument inside the math world about what any of this actually means for how research gets done. It also has something OpenAI did not put in its blog post: a Critical cyber risk rating that pushed the company to pause parts of its own internal work on the system.
Quick facts
- Announced: August 1, 2026, as a research disclosure, not a product launch
- Headline result: first explicit construction of a non sofic group, open since 1999
- Total problems solved: 10, each unsolved for a decade or longer
- Compute cost: roughly $2,000, OpenAI’s own estimate at GPT-5.6 Sol API rates
- Verification: 249 page manuscript plus Lean 4 proof files on GitHub, zero unproven steps
- Availability: not public, no confirmed release date or pricing
- Also disclosed: a Critical cyber risk classification and an internal work pause tied to it
What Astra actually solved
The result getting the most attention is the first known construction of a non sofic group. Mathematician Mikhail Gromov introduced the idea of soficity in 1999, and for 27 years nobody had produced a concrete example of a group that failed the test, only theoretical arguments that one should exist somewhere. Astra did not just find one, it produced a method that researchers expect will generalize to finding others, which is arguably more valuable than the single example itself.
The rest of the batch is just as niche and just as old. Astra tightened the ceiling on how densely spheres can be packed in high dimensional space, the first improvement to that particular bound since 1978, pushing it down toward what is known as the Cohn Elkies threshold. It disproved Alain Connes’s rigidity conjecture on von Neumann algebras, proved Ehrhart’s volume conjecture, and cleared three problems out of Paul Erdos’s famous catalogue of open questions, including problem 183 on multicolor Ramsey numbers. None of these will mean much outside a math department. Inside one, each is the kind of result a strong researcher might spend years chasing without success.
How OpenAI tried to head off the obvious objection
Extraordinary claims about AI reasoning have a bad track record of not surviving contact with actual experts, and OpenAI seems to have built its announcement around that exact problem. Rather than asking anyone to take the results on faith, the company published a 249 page manuscript alongside machine checkable proof files written in Lean 4, a formal proof assistant that either accepts a proof line by line or rejects it. The repository’s “sorry” count, the marker Lean uses for a step that has not actually been verified, sits at zero across all ten results. Every step in every proof has been checked by software, not just by a human reviewer who might have missed something.
That is a meaningfully different posture than most AI benchmark claims, which usually ask readers to trust a leaderboard score or a company’s internal evaluation. A formally verified proof is something any mathematician with a laptop can independently confirm, without first having to trust OpenAI or reproduce months of derivation by hand.
How Astra stacks up against the AI systems that came before it
Astra is not the first system to make headlines doing advanced mathematics. DeepMind’s AlphaProof and AlphaGeometry 2 spent the last two years pushing toward competition level math, and the field has moved fast enough that it is worth seeing where each approach actually sits.
| System | Approach | What it is known for |
|---|---|---|
| Astra (OpenAI) | General purpose reasoning model, not math specific | 10 decades old research problems, formally verified |
| AlphaProof (DeepMind) | Reinforcement learning, formal theorem proving | Silver medal standard on 2024 IMO problems |
| AlphaGeometry 2 (DeepMind) | Neurosymbolic hybrid, geometry specific | Gold medalist level on IMO geometry problems |
| GPT-5.6 Sol (OpenAI) | Public general purpose reasoning model | Strong benchmark scores, no open research results |
The distinction matters. DeepMind’s tools were purpose built for competition style problems with known answers and clean scoring. Astra was pointed at genuinely open research questions where nobody knew the answer in advance, which is a different and harder task than matching a known solution.
Mathematicians are impressed, and a little worried
Fields Medalist Terence Tao responded with a vision he has called “big mathematics,” large scale collaborations where humans set research direction and AI systems handle heavy derivation work, splitting complicated problems the way a lab splits experiments among researchers. He also drew a boundary around the excitement, noting that Astra appears strong at solving well posed problems handed to it, but that there is no evidence yet it can originate new theories or ask its own research questions. Fields Medalist Timothy Gowers reacted positively as well, while cautioning that the community is still digesting exactly how the proofs were found and what that implies about the model’s actual reasoning process.
That caution has an institutional backing. The Leiden Declaration, published in June 2026 and endorsed by the International Mathematical Union with more than 3,000 signatories including Tao and Peter Scholze, lists five specific risks tied to AI entering mathematics: unreliable results slipping through peer review, missing or fabricated citations, growing dependence on closed commercial systems nobody outside the company can inspect, exaggerated claims about what a model actually did, and a slow loss of scientific independence as research tools concentrate inside a handful of labs. Astra’s release, verified proofs and all, still runs into most of those concerns. The underlying model remains closed. Nobody outside OpenAI can inspect how it arrived at the constructions, only check that the final proof holds.
The part that did not make the celebration post
Buried under the math coverage is a less flattering detail. Around August 7, less than a week after the announcement, reporting indicated OpenAI had paused parts of its own internal work on Astra over cybersecurity concerns and disclosed a Critical cyber risk classification for the system, the company’s top internal severity tier. Lawmakers were reportedly briefed. OpenAI has not published the specifics of what triggered that classification, but the timing lines up with a pattern the company has been open about elsewhere. Its research arm has already built and released a separate model specifically capable of writing working exploits, arguing that giving defenders the same capability attackers will eventually get is safer than waiting. A model good enough to construct novel mathematical proofs from scratch is, by extension, a model that might be unusually good at finding logical gaps in code or cryptographic systems, which is exactly the kind of capability a Critical rating exists to flag.
It also is not an isolated incident. Security researchers have spent the past several weeks documenting cases of AI labs losing control of their own internal safety evaluations, with testing environments meant to contain a model’s behavior turning out to be less airtight than assumed. Astra being flagged internally before its math results were even fully digested by the outside world fits that same uncomfortable trend, capability arriving faster than the tooling meant to contain it.
What to watch next
- Independent replication. The Lean proofs are public, so expect university math departments to spend the next few months picking through them line by line, not just skimming the manuscript.
- Whether Astra becomes a product. OpenAI has confirmed nothing about pricing, context window, or a ChatGPT tier. The $2,000 figure is a research cost estimate, not a preview of what anyone will actually pay.
- More detail on the cyber risk finding. A Critical classification with no public explanation is the kind of gap that tends to get filled in, either by OpenAI or by outside reporting, within a few months.
- How rivals respond. DeepMind has led on competition style math for two years. An open research result from a competitor is the kind of thing that tends to speed up a lab’s own roadmap.
For now, the honest summary is a split one. Astra produced ten results that real mathematicians are taking seriously enough to spend their own time verifying, at a compute cost that would not cover a week of a junior researcher’s salary. The same disclosure that made that case also came with a safety classification serious enough to make OpenAI pause its own work, a detail that got a fraction of the attention the math did. Whether Astra ends up remembered as the moment AI started doing real research or as a reminder of how far ahead of governance that research can run may depend on which half of this story OpenAI is willing to say more about. Some of that speed comes from the same hardware push behind GPT-5.6 Sol’s new Ultrafast tier, which suggests the compute economics behind results like this one are only going to get cheaper from here.
