OpenAI says an internal version of its next model, Astra, has solved ten open problems in mathematics and theoretical computer science. Claims like that are usually impossible to check from the outside. This one is different — and that is the part worth paying attention to.

Rather than simply announcing results, OpenAI published machine-checkable proofs written in Lean, a formal proof language, on GitHub, alongside a 249-page manuscript. Anyone can download the files and run them. Lean's verification kernel gives a binary answer: the proof compiles, or it does not. There is no room for a persuasive-sounding argument that quietly skips a step.

What was actually solved

The headline result is the first explicit construction of a non-sofic group — a question that had been open since 1999. Sofic groups are a class of mathematical objects that can be approximated by finite structures; whether any group exists that cannot be approximated this way had resisted proof for over two decades.

The announcement, made on 1 August 2026, also includes new results on sphere-packing bounds and other problems across mathematics and theoretical computer science. OpenAI put the total compute cost of finding all ten proofs at roughly $2,000.

Why formal verification changes the argument

AI systems have a well-documented habit of producing confident, fluent output that is subtly wrong. In mathematics, that failure mode is fatal: a proof with one invalid step is not a partial proof, it is not a proof at all. Peer review catches these errors, but slowly and imperfectly.

Formal proof languages remove the ambiguity. Writing a proof in Lean means expressing every step in a form a computer can check against the axioms. If the model hallucinates a step, the file simply fails to compile. This is why the release matters more than a benchmark score: the claim does not depend on trusting OpenAI's description of its own work.

It is a sharp contrast with the usual difficulty of evaluating model claims, a problem we have written about in the context of verifying AI-generated content — where, in most domains, no equivalent automatic check exists.

How mathematicians responded

Reaction from the mathematical community has been positive but measured. Fields Medallist Timothy Gowers said he would recommend one of the results for publication in Annals of Mathematics — one of the discipline's most selective journals — without hesitation.

Other mathematicians have welcomed the work while noting that the results are still being digested. That caution is appropriate and normal. A Lean proof being valid is not the same as a result being important, and assessing significance takes time and human judgement. Verification tells you the proof is correct; it does not tell you the problem was worth solving.

Background: why this is not the same as past claims

AI systems have been solving competition-level mathematics for a while now. Those problems, however difficult, have known answers and are designed to be solvable. Open research problems are a different category: nobody knows whether a solution exists, let alone what it looks like.

The other important distinction is cost. At roughly $2,000 of compute for ten proofs, this is not an unrepeatable stunt requiring a national laboratory. That places it within reach of ordinary research budgets — which is precisely what makes it consequential.

What to remain careful about

Several caveats deserve stating plainly:

  • Astra is unreleased. These results come from an internal version, and internal performance does not always survive contact with a public product.
  • Selection matters. We know about ten successes. We do not know how many problems were attempted.
  • Verified is not important. Lean confirms correctness, not mathematical significance. That assessment is still underway.
  • Human framing was involved. Choosing which problems to attempt and formalising them is itself expert work.

Why it matters

If AI systems can reliably produce verifiable proofs of genuinely open problems, mathematics gains a new kind of research instrument — one that proposes candidate proofs which humans then judge for meaning. It also sets a standard the rest of the field would benefit from copying. In most AI applications there is no Lean kernel to appeal to, which is why so many capability claims stay contested. Here, for once, the evidence is checkable by anyone who cares to look.

Key takeaways

  • OpenAI announced on 1 August 2026 that an internal version of Astra solved ten open problems in mathematics and theoretical computer science.
  • The proofs were released as machine-checkable Lean files on GitHub with a 249-page manuscript, so anyone can verify them independently.
  • Headline result: the first explicit construction of a non-sofic group, open since 1999.
  • Fields Medallist Timothy Gowers said he would recommend one result to Annals of Mathematics without hesitation.
  • Total compute cost was around $2,000 — cheap enough to be repeatable, not a one-off spectacle.

The bottom line

The mathematics will take months to fully assess. The methodological point lands immediately: OpenAI made a claim that does not require anyone to trust OpenAI. In an industry where capability announcements are usually impossible to independently check, publishing proofs that either compile or fail is the most credible thing a lab can do.