For roughly the price of a used car, an artificial intelligence appears to have moved the frontier of pure mathematics. On August 1, 2026, OpenAI announced that an internal version of its next major model, code-named Astra, had produced ten new results across mathematics and theoretical computer science, each shipped with a machine-checkable Lean 4 proof certificate posted to GitHub. The total compute bill, the company said, came to about $2,000 at its published API rates.
The claim would be easy to dismiss as another lab press release were it not for two things: the problems are genuinely hard, and the proofs can be checked by anyone with a laptop. OpenAI published a 249-page technical manuscript, a separate account of how the arguments came together, and a public repository of Lean certificates under an open license. Where a human referee might take months, Lean's trusted kernel returns a binary verdict in minutes. The proof compiles, or it does not.
What Astra actually solved
The headline result is the first explicit construction of a non-sofic group, resolving a question that had stood since the mathematician Mikhail Gromov introduced the concept of soficity in 1999. Nearly every group mathematicians use in daily practice is sofic, meaning it can be approximated by finite permutation systems; whether every countable group must be was among the most prominent open questions in group theory for 27 years. Astra's answer is no.
The other nine results range widely. The model disproved the Connes rigidity conjecture on von Neumann algebras, posed by Fields Medalist Alain Connes in 1980; produced the first improvement to the general upper bound on high-dimensional sphere-packing density since 1978, pushing toward the Cohn-Elkies threshold; and resolved three problems from Paul Erdős's famous catalogue, including problem 183 on multicolored Ramsey numbers. It also proved a parallel repetition theorem for two-player quantum games and established new lower bounds on the circuit complexity of computing the permanent. Earlier, in May, the same long-horizon model family had disproved the 80-year-old Erdős unit distance conjecture in discrete geometry.
Sébastien Bubeck, OpenAI's head of mathematics research, confirmed the results on X, calling each one "beautiful" and noting that every proof arrives with a Lean certificate and a chain-of-thought walkthrough. Noam Brown, one of the researchers behind the test-time reasoning that underpins Astra, was more measured. "Sadly, no Millennium Prize Problems (yet)," he wrote, before adding a line that may matter more than any single proof: "We didn't spend a lot on each problem. It's possible to push test-time compute much further."
Perhaps the most telling endorsement came from the community's own skeptics. Thomas Bloom, the University of Manchester mathematician who curates the Erdős problems catalogue and who publicly demolished a false OpenAI math claim in October 2025, called the Astra results "big news." Fields Medalist Timothy Gowers, assessing the earlier unit distance work, said he would recommend that proof for publication in the Annals of Mathematics "without hesitation." Others who examined the new work reportedly include Noga Alon, Arul Shankar, and Jacob Tsimerman.
Why this matters
The significance is less about any one theorem than about the shape of the announcement. The bottleneck in AI-generated mathematics has always been trust. When OpenAI's model disproved the unit distance conjecture in May, validation depended on a handful of expert mathematicians reading and co-signing the argument, a strong signal but a social one that cannot be reproduced without the same rare expertise. A Lean certificate removes that dependency. It can be verified against mathlib, the community-built library of more than 210,000 formalized theorems, by anyone who runs the compiler.
The cost figure reframes the economics just as sharply. Two thousand dollars is not a research budget; it is a rounding error. But the number deserves an asterisk. It counts only the token cost of the runs that worked, at list price. It excludes failed attempts, parallel search, and the internal compute consumed along the way, so OpenAI's true outlay is almost certainly far higher. Brown's own comment that the lab "didn't spend a lot on each problem" cuts both ways: it hints at headroom, but it also concedes that the reported figure is a floor, not a ceiling.
Analysts have urged calibration. Writers at Understanding AI noted that the successes cluster in areas well suited to AI's strengths, where enough existing theory offers room to maneuver, and that none of this amounts to artificial general intelligence. No Millennium Prize Problem fell. A Lean build also proves only that the proof matches the theorem as formally stated; whether that formal statement faithfully captures the open problem as mathematicians understand it remains a human judgment.
What to watch
The immediate test is whether the broader mathematical community accepts a result announced through a blog post rather than a journal, a tension the June 2026 Leiden Declaration was written to address. Watch for independent groups to run the Lean certificates and scrutinize the formal statements, for whether OpenAI submits any proof to the Annals or a comparable venue, and for a release date and identity for Astra itself, which remains unpublished and undated. If the compute cost really can be pushed much further, the more consequential question is not what Astra solved for $2,000, but what its successor solves for $2 million.
"We didn't spend a lot on each problem. It's possible to push test-time compute much further."— Noam Brown, Researcher, OpenAI