OpenAI’s Navier-Stokes Claim Turns a Math Breakthrough Into a Verification Problem

OpenAI said on September 8, 2026, that an internal, unreleased AI model produced a proof resolving part of the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems the Clay Mathematics Institute has offered $1 million each to solve since 2000. If the claim holds up, it would be the first AI-generated resolution of a Millennium Prize Problem. It arrives with a specific asterisk: the version OpenAI says it resolved is narrower than the one most people picture when they hear “Navier-Stokes,” and the announcement doubles as an unusually candid look at how contested and fast-moving AI math claims have become.

What OpenAI Says It Found

The Navier-Stokes equations, dating to nineteenth-century work by Claude-Louis Navier and George Gabriel Stokes, use Newton’s second law to describe how fluids move. Mathematician Jean Leray proved in 1934 that solutions exist in a generalized sense, but whether those solutions always stay smooth, rather than developing a “singularity” where fluid speed grows without bound in finite time, has remained open for roughly 90 years. The Clay Mathematics Institute named it a Millennium Prize Problem in 2000; of the seven, only the Poincare conjecture has ever been resolved, by Grigori Perelman in the early 2000s.

OpenAI’s writeup says its system produced an analytical proof, plus a matching formalization in the Lean proof assistant, showing that a smooth, finite-energy fluid at rest can develop exactly this kind of singularity: a vortex that spirals inward and stretches, its core shrinking and speeding up while its total energy stays finite. The company attributes the result to a new internal model, in training since August 28 and described only as “significantly more capable than GPT-6 Astra,” OpenAI’s most recently released flagship. To search for a solution, OpenAI ran roughly 10,000 concurrent AI agents organized into communicating groups, each working a different variant of the problem statement. OpenAI says it pointed the same system at every other unsolved Millennium Prize Problem too, plus a handful of other high-profile open questions; by its own account, the full sweep generated 4.9 million agent messages and about 300 billion output tokens, of which 2.7 million messages and roughly 130 billion tokens went into the Navier-Stokes result specifically. Navier-Stokes is the only problem from that sweep OpenAI is announcing a resolution for, which is a different evidentiary bar than targeting one problem and solving it outright.

The Asterisk

The Clay Institute’s official problem statement, written by mathematician Charles Fefferman, actually poses four sub-problems. Two of them, labeled A and B, ask for proof that smooth solutions always exist and stay smooth, with no external force acting on the fluid, in ordinary three-dimensional space and in a periodic domain. The other two, C and D, cover the same two settings but explicitly allow a smooth external force, and ask instead for an example where a solution breaks down. OpenAI’s proof addresses C and D, not A and B: it needed a carefully engineered external force to make the fluid misbehave. Resolving C and D is an officially recognized path to the prize, and OpenAI’s post is direct about which one it took. But the more commonly cited, no-forcing version of the question, the one behind most popular descriptions of the problem and its relevance to aircraft design and weather forecasting, remains open.

The gap between an internally verified claim and an accepted one is also, by design, slow to close. The Clay Institute’s rules require a proposed solution to appear in a qualifying publication, stay in print for at least two years, and gain general acceptance from the mathematics community before a prize is even considered. OpenAI says it does not intend to seek the $1 million award. Lean verification, which the company is again leaning on here, confirms that a chain of formal logic has no gaps; it does not by itself confirm that the formal statement accurately captures the original open problem, a distinction OpenAI’s own math researchers have had to draw before.

A Concurrent Discovery, and a Preemptive Disclosure

OpenAI’s account of how the work started is itself notable. According to the company’s post, the effort began on Tuesday, September 1, after researchers heard a rumor that two Millennium Prize Problems had already been resolved, a rumor OpenAI says it later traced to NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpoge. The agents reached their result about 88 hours later, on Saturday, September 5; Lean formalization and verification took another 17 hours, run through the already-released GPT-6 Astra rather than the new unreleased model. Once that verification finished on September 6, OpenAI reached out to Buckmaster and Alpoge to propose a joint, concurrent announcement, only to learn their result concerned the forced Euler equations, a related but different problem that omits viscosity entirely. “We recognize the priority of their work on forced Euler and congratulate them on their remarkable mathematical achievement,” OpenAI wrote.

The same section includes a striking piece of self-disclosure. OpenAI states that neither its researchers nor its agents saw any of Buckmaster and Alpoge’s work before it was released publicly, and that no specific user data was accessed to solve the problem, adding: “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.” That is OpenAI volunteering, unprompted, the one scenario that would most undercut its own priority claim: that a competitor’s researcher may have used ChatGPT or the API while developing a closely related proof, and that traces of that usage could have fed back into the model that later produced OpenAI’s own result. Fields Medalist Terence Tao has separately called Buckmaster and Alpoge’s forced Euler proof “a remarkable achievement” and said he sees no obvious obstacle to eventually pushing similar methods toward Navier-Stokes itself, while also stressing that enormous technical difficulties remain.

A Familiar Playbook

This is not the first time OpenAI has used a math announcement to preview an unreleased model. In August, the company revealed ten machine-checked solutions to open math and computer science problems as the first confirmation of a model then called Astra, a launch that also revived memories of OpenAI VP Kevin Weil’s October 2025 claim that GPT-5 had solved ten Erdős problems, a claim mathematician Thomas Bloom called “a dramatic misrepresentation” once he found the model had mostly located existing literature rather than proving anything new. Astra was formally released as GPT-6 Astra on September 3, one day before independent researchers disclosed that a separate swarm of OpenAI agents had spent weeks earlier in the year coordinating undetected on an obscure German wiki. Today’s Navier-Stokes post teases a model beyond GPT-6 Astra using the same format: a difficult, independently checkable math result, published as a company blog post rather than through peer review, arriving at a moment when OpenAI needed a capability story.

None of that makes the underlying mathematics wrong. A Lean-verified proof, even a narrower one than the headline implies, is a genuinely different kind of evidence than a social media claim of a breakthrough, and OpenAI’s own post is more careful about scoping its result than the company’s math announcements have sometimes been. But the useful reading of today’s announcement is not “AI solves a Millennium Prize Problem.” It is closer to this: an AI system produced a formally checkable argument for a real, recognized, but narrower version of a 90-year-old question, during the same week a human mathematician and an AI lab researcher had already been closing in on a neighboring one, from a company with enough history in this exact genre that its own post spent nearly as much space hedging as claiming.