OpenAI Says It Solved Navier–Stokes: Why It Is Not Final Yet, but Could Still Be a Turning Point for Mathematics

The Navier–Stokes problem, one of the seven famous Millennium Prize Problems, may have received the first proof created by AI. OpenAI published both a mathematical paper and a Lean formalization produced by a system of thousands of agents. But official recognition will require years of independent scrutiny.

There are statements in mathematics so difficult that for decades they serve as markers of the boundary of human knowledge. On Sept. 8, OpenAI said one such boundary may have been crossed: its experimental artificial intelligence system had constructed a proof for the celebrated Navier–Stokes existence and smoothness problem.

This is not simply another difficult benchmark for a language model. Navier–Stokes is one of the seven Millennium Prize Problems designated by the Clay Mathematics Institute in 2000. Each carries a $1 million prize, and in more than a quarter-century only the Poincaré conjecture has been officially resolved.

OpenAI says its system found a specific scenario in which the smooth motion of a three-dimensional incompressible fluid governed by the Navier–Stokes equations can develop a singularity in finite time — a mathematical regime in which velocity ceases to remain bounded.

That does not mean water in the physical world can suddenly acquire infinite speed or “explode” because the laws of physics fail. If the proof is correct, it instead exposes a limit of the mathematical model itself. Under certain conditions, the equations would cease to describe a physical fluid adequately as a continuous medium.

Daycom’s analysis indicates that the real sensation is not merely that a Millennium Problem may have been solved. More important is who — or what — produced the proof. If it survives scrutiny, one of mathematics’ most prestigious summits may for the first time have been reached by an AI system rather than by an individual mathematician or a human research team.

The Navier–Stokes equations describe the motion of fluids and gases. They are used in models of airflow around aircraft, weather forecasting, blood flow and many engineering applications. Yet the fundamental question of how their three-dimensional solutions behave has remained open for nearly a century.

In simplified terms, the problem asks this: if a fluid begins in a perfectly “well-behaved” state, is its motion guaranteed to remain mathematically smooth forever? Or can the equations generate a point at which certain quantities become unbounded and the classical solution effectively breaks down?

The official Millennium Problem formulation allows two broad routes. One can prove that smooth solutions always exist, or construct a valid example in which they break down. In formulations C and D, a smooth external force is explicitly allowed so long as it satisfies the prescribed conditions.

According to OpenAI, its system pursued the second route. It begins with a fluid in a smooth state and constructs a vortex structure that becomes increasingly compressed and stretched. Its central region shrinks, its rotation accelerates, yet the total energy remains finite.

The critical point is that the singularity is not created artificially by inserting an infinite external force. In the researchers’ description, the relevant terms in the equations grow while canceling one another with extraordinary precision, allowing the forcing to remain smooth even as the velocity becomes unbounded.

If that construction is mathematically flawless, it would satisfy precisely the alternative contemplated by the Clay Institute’s original formulation: identify admissible initial conditions and a smooth force for which no global smooth solution exists.

But the word “if” carries enormous weight here. OpenAI has announced a solution; the mathematical community has not declared the problem closed. As of Sept. 9, the Clay Mathematics Institute still lists Navier–Stokes among the unsolved Millennium Prize Problems.

That is not a formality, nor is it distrust directed specifically at artificial intelligence. It is how mathematics works. A complicated proof can contain one almost invisible flaw capable of invalidating dozens of pages of otherwise correct reasoning. The history of the field is full of claimed solutions to great problems that failed independent review.

The Millennium Prize rules deliberately set an exceptionally high bar. The Clay Mathematics Institute does not accept a proposed solution directly from its authors for adjudication. The work must appear in a qualifying mathematical publication, at least two years must pass after publication, and the proof must gain general acceptance within the international mathematical community.

That means even a perfect proof released this week cannot turn the label “unsolved” into “solved” tomorrow. In the case of the Poincaré conjecture, years also passed between Grigori Perelman’s 2002–2003 preprints and the formal award of the Millennium Prize in 2010.

OpenAI, however, has said it does not intend to claim the million-dollar prize. The company presents the project primarily as a demonstration of a new level of capability and has released not only a conventional mathematical paper but also a formalized version of the proof in Lean.

Lean substantially changes the nature of verification. In an ordinary mathematics paper, experts read the argument and check the logical transitions between statements. A formal proof translates that logic into a form that can be mechanically checked against a defined set of axioms and inference rules.

That does not make the proof automatically correct in the broader sense. It still matters whether the original problem was formalized accurately, whether the definitions correspond to the intended statement and whether the mathematical bridge between the formal construction and the claimed result is sound. But the space for ordinary human error becomes significantly narrower.

Just as striking is the way OpenAI says it searched for the solution. The company did not place a single model in front of the equations and ask it to “think longer.” It deployed a system of parallel agents divided into groups, with different teams exploring different approaches and exchanging promising results.

The group that ultimately reached the Navier–Stokes result consisted of roughly 10,000 agents operating simultaneously. The first agents started on Sept. 1, and the result was obtained on Sept. 5 — after about 88 hours. Formalization and verification in Lean took roughly another 17 hours.

The computational scale was unusual even by modern AI standards. During the Navier–Stokes effort alone, the agents exchanged about 2.7 million messages and generated roughly 130 billion output tokens. Across the broader series of mathematics experiments, about 300 billion tokens were used.

That matters because it changes the familiar image of an “artificial mathematician.” This is not a digital equivalent of a lone genius working over a sheet of paper. The architecture increasingly resembles a vast research institute in which thousands of agents test ideas, abandon dead ends and pass the most promising lines of reasoning onward.

Humans did not disappear from the process. Researchers directed the groups, transferred key ideas between them and, after an intermediate success, shifted more resources toward Navier–Stokes. This is therefore not a story about a completely autonomous AI that independently woke up one morning having solved a Millennium Problem.

The initial push came from another result. The system first worked on a simpler, related problem involving the Euler equations — effectively a version of fluid-motion equations without viscosity. After success there, the team concluded that Navier–Stokes looked like the most promising direction.

A dispute over scientific priority has already emerged around parallel work in the field. OpenAI said it intensified its effort after hearing reports of results by mathematicians Levent Alpoge and Tristan Buckmaster, which turned out to concern a different, though related, problem involving the Euler equations.

The company says its researchers and agents did not see their unpublished work and did not gain access to specific user data in order to solve the problem. At the same time, it acknowledges that it cannot completely rule out the influence of anonymized product-usage data on earlier improvements to its models.

That dispute may foreshadow a new challenge for science. When AI can absorb enormous quantities of information and develop ideas through thousands of parallel processes, it becomes harder to determine where general model training ends and where the specific intellectual contribution of an individual researcher begins.

An even more fundamental question concerns the nature of mathematics itself. Historically, proving a great theorem was never only about the final answer. Decades of failed attempts often generated new methods, concepts and entire areas of research. The path to the solution could be more valuable to science than the solution itself.

If a machine can explore in 88 hours an intellectual space that would take a human community decades to navigate, mathematics may move from a scarcity of proofs to an abundance of them. The central challenge would then become not only finding truth, but understanding which among millions of machine-discovered routes actually give humans a new picture of the subject.

That is where the paradox emerges. A formally proved theorem can be entirely correct and still be poorly understood by people. AI may be able to find a path through the mathematical landscape, while human science must still explain why that path works, which ideas within it are universal and where they lead next.

Until now, major AI results in mathematics could often be described as exceptionally powerful extensions of human ideas. A system might find a shorter proof, combine known techniques or push the boundary of an earlier result. Navier–Stokes, if the proof survives, would change the scale of comparison.

The question would no longer be simply, “Can AI help great mathematicians?” It would become far more uncomfortable: can a system itself produce results of historic importance faster than human mathematical culture can verify them, understand them and absorb them into its own body of knowledge?

In the coming months, the answer will depend not on OpenAI presentations, but on mathematicians taking the proof apart line by line. They will examine the analytic estimates, the singularity construction, compliance with the exact conditions of the problem and the formal Lean version. That process will determine the result’s status.

So for now, the accurate formulation is not “AI solved the Navier–Stokes problem” as a settled historical fact, but “OpenAI published a candidate solution.” The Clay Mathematics Institute still lists the problem as open, and the prize rules contemplate years before possible formal recognition.

Even that cautious status does not make the event small. Thousands of agents have already produced a mathematical object serious enough to place one of analysis’s oldest open problems under immediate global scrutiny.

If a fatal flaw is found, the experiment will still show how close AI has moved to the frontier of pure mathematics. If no such flaw is found, history will remember more than the closure of a second Millennium Prize Problem.

It will remember the moment when the question about artificial intelligence in science changed.

Not “Can a machine help us find a proof?”

But “What will mathematicians do when machines begin finding the proofs?”