The files appeared on GitHub on Monday evening: 722 manuscripts, arranged in 372 groups, each built around a single claimed mathematical result. They were produced by an OpenAI model that the company has not released to the public, and they touch problems that have resisted mathematicians for decades, including, by the company’s account, three of the five remaining Millennium Prize Problems.
The release, in a repository called openai/math under an Apache-2.0 license, came roughly a month after OpenAI claimed an internal model had solved the Navier-Stokes existence and smoothness problem, one of the seven problems the Clay Mathematics Institute has backed with a $1 million bounty each. That earlier claim, a 165-page proof assembled with thousands of AI agents working in parallel, drew sharp criticism from mathematicians who said the company had not shown enough for others to verify.
The new drop is far larger. The papers span 17 areas of mathematics, from number theory to algebraic geometry to quantum physics. OpenAI said it fed the model about 4,000 problems in total and that the release includes Lean formalizations of many proofs, the programming language that checks each logical step, along with protocols for revising papers and citing the work.
Chief executive Sam Altman has singled out four results: the quasi-Riemann hypothesis, the Unique Games Conjecture, a case of the Hodge conjecture, and the free group factor problem. The company is careful about the framing. It does not claim to have proved the Riemann hypothesis itself, only a weaker statement about where the zeta function’s zeros can sit. The Hodge result, likewise, covers one family of shapes, not the conjecture as a whole.
That gap between the headline and the actual claim is where the mathematical community’s skepticism has concentrated. None of the 722 manuscripts has been peer-reviewed, the process by which independent experts check a proof before a journal publishes it. The model that produced them remains internal, so no one outside OpenAI can rerun the experiment. Several mathematicians, still smarting from the Navier-Stokes episode, have asked the company for the full reasoning, not just the conclusions.
The release also lands in the middle of a broader debate about whether AI is finishing mathematics or merely automating it. Weeks ago, the company told The New York Times it had made “substantial progress” on a second Millennium Prize Problem without naming it. Believers and critics alike have argued over whether a machine that works through thousands of problems has done mathematics, or only something that looks like it.
The Wall Street Journal has sounded a different caution, one aimed at investors rather than mathematicians. The paper noted that breakthroughs of this kind do not translate directly into product revenue. A proof of a Millennium-class statement, even if it survives scrutiny, does not by itself sell a chatbot subscription or win a cloud contract.
Altman has framed the moment in grander terms, calling it the start of an era of discovery. The company plans to fund workshops and conferences around the findings and says it is working to release the model behind them. For now, that model stays behind closed doors, which leaves the field with a familiar standoff: a company insisting it has made a leap, and a discipline that has spent centuries insisting that a leap counts only once it can be checked.
The economics sit behind the announcement too. OpenAI is competing for the same pool of AI infrastructure capital that companies like SpaceX and Anthropic are borrowing to fund, and a demonstrated capability at the hardest problems science has to offer is a form of evidence that the spending is not wasted. Whether that evidence holds up is now a question for the mathematicians, not the marketing team.
The reception has split along predictable lines. Some researchers see the release as the first serious challenge to the assumption that machines cannot do original mathematics, and have begun the slow work of checking the proofs. Others point out that the most consequential claims, the ones touching the Riemann hypothesis and the Hodge conjecture, are exactly the ones the company has framed most narrowly, and that the full model remains out of reach. The truth will emerge, as it always does in mathematics, one checked line at a time.
For OpenAI, the release does something no amount of benchmark scores could: it places the company’s internal model inside the oldest and strictest verification culture in science. The math will be checked line by line by people who have no stake in the company’s valuation. That is precisely what makes the release a gamble as much as an announcement. The next few months, spent in the unforgiving light of peer review, will decide whether the trove is a breakthrough or a bluff.
