Talk Around Town: OpenAI Claims 722 AI Math Manuscripts Solve 372 Open Problems, but Mathematicians Ain't Buying It Yet

What the Chatter Says Happened

Well, slap the pig and call it Sunday: on October 6 and 7, 2026, OpenAI posted a GitHub repository containing 722 mathematical manuscripts, according to independent reporting by The Washington Post, Fortune, Scientific American, Engadget, and The Decoder. The company says those manuscripts cover 372 families of previously unsolved research problems, all generated by an internal AI model that OpenAI has not publicly named.

OpenAI says most of these results came out of a single prompt fed to a single AI agent, and the company claims the model averaged roughly three hours of compute time per solution, according to Engadget and Fortune. That is a pace of mathematical output that would make a tenure committee weep into their coffee — or, depending on who you ask, run screaming from the building.

What Is Actually Known and Confirmed

The publication of the GitHub repository itself is confirmed by multiple independent top-tier outlets. This October release follows a September 21, 2026 OpenAI announcement in which the company claimed its internal model had resolved more than 100 long-standing open problems, including a claimed resolution of the Navier-Stokes Millennium Prize Problem, according to Tech Insider and ExplainX.

OpenAI also confirmed it formed an Advisory Group on Mathematics and Artificial Intelligence, hosted at the Institute for Advanced Study, which includes prominent mathematicians such as Edward Witten and Timothy Gowers, according to ExplainX and Tech Insider. However, OpenAI confirmed it drew one hard line before that group ever sat down: according to The Decoder, the advisory mathematicians can weigh in on how results get communicated to the public, but the company retained full control over whether and how fast the AI keeps producing more of them. That is a bit like hiring a restaurant critic to advise on how your barbecue gets plated while telling them they have no say in whether the hog is cooked through.

The Decoder and Tech Insider both report that roughly half of the 372 result families lack Lean formal proofs, meaning machine verification is not yet available for a substantial portion of the claimed results. That gap has not been confirmed by any peer-reviewed audit as of this writing.

What Prominent Mathematicians Are Actually Saying

Fields Medalist Terence Tao did not mince words, publicly calling the pace of AI-generated mathematical output 'insane,' according to Fortune, Scientific American, and The Decoder. Tao argued, per those outlets, that AI is churning through problems in a fashion that is not sustainable and may actively discourage students from entering mathematics — which is a hell of a thing to contemplate when you remember that every one of those 372 problems once represented somebody's life's work.

NYU mathematician Tristan Buckmaster raised a practical objection that cuts to the bone: according to Tech Insider, Buckmaster argued it simply is not plausible that 372 result families could be thoroughly reviewed by qualified experts in the window between generation and public release. Think of it as a neighbor dropping seven hundred and twenty-two casseroles on your porch at midnight and telling you they're all fine to eat — you'd want somebody to check at least a few of them first.

Timothy Gowers, who sits on OpenAI's own advisory group, offered a warning that goes beyond the immediate release. According to The Decoder, Gowers cautioned that within a decade or two, mathematical literature could expand enormously while no human community remains that genuinely understands it. That is the kind of observation that keeps a person staring at the ceiling fan at three in the morning.

University of Toronto mathematician Daniel Litt offered a different take, arguing according to Scientific American that open publication is preferable to secrecy and that the results will ultimately benefit mathematics. So not every voice in the field is hollering from the same stump, and that disagreement matters.

What Remains Unverified and Murky as a Catfish Pond

One outlet, BigGo Finance, reported that scholars accused OpenAI of ignoring requests made at a closed-door August 2026 meeting to follow standard academic publication norms, and that multiple mathematicians reportedly described the company's conduct as 'mafia-like.' That is a colorful allegation, but it has not been independently corroborated across multiple outlets consulted for this article, so it remains an attributed allegation only.

BigGo Finance also reported an allegation that OpenAI scooped a collaborative Navier-Stokes effort that involved an NYU mathematician and an Anthropic employee. That claim likewise has not been independently confirmed by other outlets at the time of writing, and should be treated with appropriate skepticism until it is.

OpenAI has not publicly disclosed which internal model generated the manuscripts, what compute budget was allocated per problem, or what internal standard of proof was required before a result was included in the release, according to the available reporting. Those gaps make independent assessment about as easy as trying to judge a pie-eating contest you weren't invited to.

Analysis: The Barn Is Full, but Has Anyone Checked the Roof?

This is analysis, not reporting. The sheer volume of the October release — 722 manuscripts in roughly two days — is not inherently evidence of correctness or incorrectness. Mathematics has always progressed in bursts, and machine-assisted proof generation has been a legitimate area of research for years. But volume and validity are not the same bucket, and dumping seven hundred manuscripts into the public domain before rigorous external review is a genuinely new situation for a field built on the slow, painful labor of human verification.

The advisory group arrangement, as confirmed by The Decoder, looks like an attempt to thread a needle that may not actually exist: you cannot have a truly independent review body if that body is explicitly excluded from influencing the pace of production. That structural limitation seems likely to make the advisory group's imprimatur less reassuring to skeptical mathematicians, not more.

The community divide — between those like Daniel Litt who see open publication as a net positive and those like Terence Tao and Timothy Gowers who fear cultural and epistemic damage — reflects a real tension that is not going to resolve itself before the next batch of manuscripts arrives. If OpenAI's claimed results hold up under scrutiny, this will be a remarkable moment in the history of science. If they don't, the field will have spent enormous time and energy chasing a herd of goats in every direction. Right now, nobody outside OpenAI knows which one it is.