OpenAI publishes 372 new math results, straining a truce with mathematicians

OpenAI said late on Oct. 6 that an internal model had produced 372 new results in mathematics and theoretical computer science, publishing the work in a public GitHub repository rather than a peer-reviewed journal. The company said nearly all of the results came from a single prompt handed to a single AI agent, a sharp contrast to the thousands of parallel AI "agents" and millions of dollars in computing time it says it spent cracking the Navier-Stokes equations just a month earlier.

The drop is the third and by far the largest escalation in a monthslong standoff between OpenAI and the pure-mathematics community over how fast AI-generated proofs should be announced, who gets credit for the ideas behind them, and who has time to check whether they are actually correct. OpenAI says the new results touch 16 areas of mathematics and include progress on the long-unsolved four-dimensional Kakeya conjecture and incremental advances relevant to the Riemann hypothesis, one of the seven Clay Mathematics Institute Millennium Prize problems. The announcement was reported in detail by Scientific American, whose account forms the core of what is confirmed so far about the release.

A much bigger number, a much smaller price tag

In August, OpenAI published what it called "Ten Advances in Mathematics and Theoretical Computer Science," a 249-page manuscript built around an internal, unreleased version of a model it has referred to as Astra. That release cost roughly $2,000 in computing time, by the company's own account, and came packaged with machine-checkable Lean proof certificates so outside mathematicians could verify the logic step by step.

Two months later, the scale has changed by more than an order of magnitude. The October release groups its findings into 372 distinct result families, and OpenAI has said the average result required computing power roughly equivalent to a few hours of its consumer "ChatGPT Pro" tier, a fraction of the resources poured into the Navier-Stokes claim. Many, but not all, of the new results ship with the same kind of Lean formalization used in August; others do not, leaving mathematicians to work through the underlying arguments by hand.

What OpenAI has not published is just as notable as what it has. An independent advisory group the company helped set up after the August release had recommended that any lab publishing AI-generated mathematics disclose the specific model used, the exact prompt given to it, and the compute time spent on each individual result. OpenAI instead released only an average compute figure across all 372 results, along with some additional aggregate statistics, and no prompts. A spokesperson noted the company is "not bound" by the advisory group's recommendations, according to Scientific American's reporting.

How a math demo turned into a standoff

The dispute did not start this week. When OpenAI published its first ten results on Aug. 1, several working mathematicians who reviewed the proofs said parts of the arguments closely tracked earlier published work without adequate citation. Steven Miller, a number theorist, argued that an improved bound on high-dimensional sphere packing leaned on a specific technique from his own 2016 paper, while Cambridge's Francesco Fournier-Facio said a separate result on group theory combined methods already published in 2016 and 2019. OpenAI said at the time that it was holding itself to "the same standards generally expected of human mathematicians" and promised revisions, a response examined in a follow-up Scientific American report in which outside experts went further and called the lapses research misconduct.

Weeks later, OpenAI said an updated internal model had made progress on the Navier-Stokes existence and smoothness problem, another Millennium Prize question, using a massive parallel search involving thousands of AI agents. The claim drew a public objection from Tristan Buckmaster, a Courant Institute professor at New York University who works on fluid-dynamics equations, who said he had been discouraged from crediting a collaborator who happens to work at a rival AI lab.

That episode helped prompt a formal response from the field's most decorated figures. On Sept. 11, more than two dozen Fields Medal winners, spanning laureates from 1978 to 2026, signed a joint statement titled "A Severe Misalignment of AI in Mathematics," arguing that AI labs' race to rack up solved problems as a capability benchmark works against how mathematics actually advances. "Often these solutions are announced in a rush, leaving no time for a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others," the signatories wrote, adding that solving a problem is "only a tool and proxy for achieving the primary goal of conceptual understanding and insight." The declaration does not call for AI tools to be withdrawn from mathematics, only for the field, not raw problem counts, to set the pace.

Mathematicians divided over how to respond

OpenAI's answer to that pressure was to help stand up a nine-member advisory group, which includes Fields medalist Timothy Gowers among mathematicians drawn largely from the Institute for Advanced Study's orbit. Its members are unpaid, have no authority over how fast OpenAI publishes and are intended, in the company's framing, to help assess results, coordinate disclosure and uphold ordinary standards of mathematical credit. The October release is the group's first real test, and reaction to it has split along familiar lines.

Andrew Sutherland, a senior research scientist in MIT's mathematics department, said the claim that nearly all 372 results came from one prompt to one AI agent cannot be taken at face value yet. "Until and unless they release the model and people can replicate their results, I think you should treat any claims about one-shotting problems with a single agent as unverified," he said, as quoted by Scientific American.

"If we want to know the answers to these math questions, I see no reason why we should ask the company to keep them secret."

That was the view of Daniel Litt, a number theorist at the University of Toronto, who argued the underlying mathematics should be shared regardless of how uncomfortable the disclosure process has been. Terence Tao, the UCLA mathematician and Fields medalist who helped organize September's declaration, has separately called the overall pace of AI-generated claims "insane," warning that even the rumor of a lab working on a famous problem can now trigger a flood of AI effort aimed at resolving it before the human researcher who started it can finish. An earlier, independent audit of OpenAI's August proofs, posted to arXiv, found that while most of the Lean-formalized arguments checked out logically, attribution and framing problems were harder to catch by machine than errors in the mathematics itself, a finding documented in the preprint circulating among reviewers.

The immediate burden falls on working mathematicians, many of them unpaid volunteers, who must now check, attribute and contextualize hundreds of machine-generated arguments on top of their existing research and teaching loads. Journals and conference organizers face a related problem: deciding whether AI-generated proofs without named human authors can be submitted, cited or credited at all under existing norms. Universities and funding bodies are watching too, since the episode bears on how research credit, tenure cases and grant applications treat machine-assisted discovery going forward. And rival AI labs, including Anthropic and Google DeepMind, are being drawn into the same disclosure debate by association, since the competitive pressure to claim mathematical milestones is not unique to OpenAI.

For OpenAI, the stakes are more commercial than they might appear. Mathematical reasoning benchmarks have become a proxy war among frontier AI labs for demonstrating raw model capability to investors and enterprise customers, making each new result both a scientific claim and a marketing one. That dual purpose is precisely what the Fields medalists' declaration objected to, and it is also why OpenAI has strong incentive to keep publishing at a pace the advisory group was created to slow down.

What happens next

The advisory group has not yet issued a public assessment of the October release, and it is unclear whether it was consulted before publication or only after. Outside mathematicians are expected to spend the coming weeks working through the 372 result families by hand and by machine, checking Lean certificates where they exist and flagging citation gaps where prior work appears to have been missed, much as happened after August. OpenAI has not said whether it will eventually release the underlying model itself, which Sutherland and others argue is the only way to settle the question of whether a single prompt really can generate hundreds of advances without extensive hidden iteration. Until then, the field's most visible figures have signaled they intend to keep pushing OpenAI toward the slower, more documented process mathematics has traditionally relied on, even as the company shows no sign of slowing down.