On September 8, 2026, OpenAI announced something that seemed almost too large to process: an internal AI system had produced a Lean-formalized proof that smooth three-dimensional Navier-Stokes flow can develop a finite-time singularity. The Navier-Stokes existence and smoothness problem is one of seven Millennium Prize Problems — a set of questions so difficult that the Clay Mathematics Institute offers $1 million for a valid solution to each. Mathematicians had been chipping away at Navier-Stokes for over a century.
OpenAI said it did not intend to claim the prize.
Then, on October 6, the company did something more provocative: it opened a GitHub repository and dropped 722 mathematical manuscripts — produced by the same unreleased internal model — into the open under an Apache-2.0 license. The repository, `openai/math`, contained results spanning 372 distinct “families,” covering topics from the irrationality exponent of π to Kaplansky’s direct-finiteness conjecture to the Mézard–Parisi formula in spin glass physics. Each result was generated by posing the model roughly 4,000 mathematical problems. The average resolution cost approximately three hours of ChatGPT Pro-equivalent compute. A few took considerably more.
The reaction from the mathematical community has been something between awe, caution, and frustration — and it is still unfolding.
What the Repository Actually Contains
The count itself has been confusing. The repository initially listed 722 manuscripts; later accounting settled at 719 distinct results. The discrepancy matters less than what the numbers represent. A “family” can include a main result, companion arguments, consequences, and alternative proofs, so 719 is not 719 independently solved research-level problems. Still, the scope is extraordinary: organized into 372 families, the collection spans fields from number theory and algebraic geometry to quantum statistical mechanics and theoretical computer science.
Ten of the results come with shortened summaries of the model’s reasoning. The topics of those summaries hint at the breadth of the release: multiplicative function correlations, arithmetic progressions, the Mahler conjectures, NP-hardness at a semidefinite programming threshold, free group factors, and the 3D relativistic Vlasov-Maxwell system. Several results — including the Hodge Conjecture for CM abelian varieties and a zero-free region improvement for the Riemann zeta function — received special attention in the announcement.
The Lean formalization picture is incomplete. About 42% of the 719 top-line results — roughly 300 manuscripts — have been formalized in Lean, a proof verification system that checks whether a stated argument follows from its premises with machine precision. That is a meaningful quality control signal. But it is not a guarantee: Lean verifies that a formalized statement follows from a formalized proof, not that the formal statement actually captures the informal mathematical claim the manuscript describes. For the remaining 58%, OpenAI’s own README warns that unformalized results “may contain issues,” and promises to add formalizations as they become available.
That warning proved prescient almost immediately. By October 7 — one day after the release — three manuscripts had been withdrawn. A sign error was found in one argument, and two papers that depended on the flawed construction were pulled along with it. Fourteen additional manuscripts were revised to repair proofs, correct statements, clarify hypotheses, or fix dependency chains.
A Timeline Built in Weeks
The speed of the whole enterprise is as remarkable as its scope. OpenAI says training of the internal model began on August 28. A focused evaluation effort targeting Millennium Prize problems launched September 1 — just four days later. The model’s agents reached the Navier-Stokes resolution on September 5. The public announcement followed on September 8. The model’s advisory group published its recommendations on September 29. The full manuscript dump landed on October 6.
From the start of training to the release of 722 papers: 39 days.
The model is not public. OpenAI has not named it, disclosed its architecture, or given a timeline for any release. What is known is that it operates in an agentic configuration — pursuing mathematical problems over extended compute budgets, with individual results sometimes requiring substantially more than the three-hour average. The Navier-Stokes proof used a different procedure from the bulk of the manuscripts; so did the Hodge Conjecture for CM abelian varieties result, which was noted separately in the release.
What Mathematicians Are Saying — and Not Saying
OpenAI consulted the Advisory Group on Mathematics and Artificial Intelligence, an independent body associated with the Institute for Advanced Study in Princeton. The group received over 600 responses from the broader mathematical community in the course of its work.
Its assessment of the release is diplomatically measured: the advisory group called the publication “an important event.” It did not endorse OpenAI’s process. It did not assess the quality or significance of the mathematical work. And in a pointed statement, it said it “does not endorse this practice” — meaning the testing of advanced mathematical problems on proprietary models — and asked OpenAI to stop.
The group’s recommendations go further. It has asked that results be deposited in repositories no AI laboratory controls; that future releases disclose the model name, prompts, chain-of-thought summaries, compute time, and cost; that proofs be formalized where possible; that comparison data be published on problems the models attempted but failed; and that funding be directed toward human understanding efforts and broad access to the models.
OpenAI’s response has been to say it is “exploring community-hosted alternatives” for future deposits. The company has so far not named the model, disclosed its full reasoning chains, or published its failure cases.
Why This Is Different From Every Other AI Benchmark
The broader AI industry has grown deeply skeptical of benchmark results. Companies release models, claim state-of-the-art performance on carefully selected evaluations, and the community waits months to understand whether the improvement is real. Contamination — training data that overlaps with the test set — has become endemic. Human evaluators sometimes disagree on what a correct answer even is.
Mathematics is different. A proof is either valid or it is not. Lean does not grade on a curve. The Clay Institute did not design Navier-Stokes as a benchmark; it is a two-century-old problem that describes how fluids behave. There is no contamination to worry about in the Hodge Conjecture.
That is why the 42% Lean verification rate matters as much as it does. It means the other 58% are, for now, claims — human-readable arguments that mathematicians will need to check themselves. Some of those will be correct and simply not yet formalized. Some may have errors. Three already did.
And it is why the advisory group’s discomfort deserves to be taken seriously. The group’s concern is not that the results are false. Its concern is structural: when an unreleased, undisclosed model produces hundreds of mathematical claims and a company publishes them under its own brand, with its own timeline and its own selective disclosures, the mathematical community loses the ability to evaluate the provenance of the work. Math has always depended on open argument. You show your work. You disclose your method. Other mathematicians check it and find the errors — as they did, in one day, with the three manuscripts that were withdrawn.
The Stakes Beyond the Proof Count
The implications of a general-purpose mathematical reasoning system extend well beyond pure mathematics. The same capabilities that allow a model to navigate spin glass physics apply to cryptographic proofs, formal verification of software, and the design of error-correcting codes. A system that can credibly advance research in number theory is close, in capability space, to a system that can find vulnerabilities in cryptographic protocols or prove the correctness of safety-critical systems.
OpenAI has not said what it plans to do with the model commercially, or when. The company did say it is not claiming the Millennium Prize — a decision that is itself interesting. The Clay Institute’s rules require that a result be published in a peer-reviewed journal and stand unchallenged in the literature for two years. An Apache-2.0 GitHub dump does not satisfy those criteria, and 39 days is not two years.
What October 6 represents is something more like a demonstration than a completion. A system that can produce 722 papers in mathematics — even imperfect ones, even with three immediate withdrawals and fourteen revisions — is operating in intellectual territory that a year ago most researchers would have described as a decade away. The math community is right to proceed carefully. The manuscripts deserve scrutiny. The process deserves transparency.
But the results also deserve to be taken seriously. Forty-two percent of 719 is 300 Lean-verified proofs, from a model that has existed for 39 days. Whatever its limitations, something is happening in that GitHub repository that mathematics has not seen before.
—
KickHammer Tech covers artificial intelligence, emerging technology, and the companies shaping the future. Have a tip? Reach us at tips@kickhammernews.com.
