OpenAI Says AI Found a Navier–Stokes Blowup Proof

On September 8, 2026, OpenAI published a claimed solution to one of mathematics’ most stubborn open questions. Its 166-page Navier–Stokes paper constructs a three-dimensional fluid flow whose velocity becomes unbounded in finite time, even though the flow starts smoothly and retains bounded kinetic energy. OpenAI also released a formal proof written in Lean.

The company says the proof came from roughly 10,000 coordinating AI agents running a next-generation model “significantly more capable” than GPT-6 Astra. The agents reached the result after about 88 hours, followed by another 17 hours of Lean formalization and verification using Astra. Training of the unnamed internal model is still underway.

If the argument survives independent scrutiny, it would settle the Navier–Stokes Millennium Prize Problem by showing that smooth solutions can break down. That “if” still matters. As of September 9, the Clay Mathematics Institute continues to list Navier–Stokes as unsolved, and its prize rules require a qualifying publication, at least two years of scrutiny, and general acceptance by the mathematics community.

OpenAI Targets the Prize’s Breakdown Route

The Navier–Stokes equations describe the motion of fluids as continuous materials. Velocity changes according to acceleration, pressure, viscosity and external forces, while an incompressibility condition prevents the fluid from being squeezed into a smaller volume without expanding elsewhere.

These equations support much of modern fluid mechanics, including models used in aircraft design, weather forecasting and blood-flow research. The unresolved mathematical issue is whether a smooth three-dimensional solution must stay smooth forever, or whether some quantity such as velocity can become infinite after a finite amount of time.

The common summary of the problem focuses on an unforced fluid evolving from smooth initial conditions. The official Millennium Prize formulation, however, defines four acceptable routes to a resolution. Alternatives A and B ask for global smoothness without an external force. Alternatives C and D allow a solver to demonstrate breakdown using smooth initial data and a smooth force, respectively on three-dimensional Euclidean space and a periodic three-dimensional torus.

OpenAI’s proof targets C and D. It constructs a fluid that begins at rest, then receives a smooth force compactly supported in space and time. The resulting velocity and pressure remain smooth before the singular time, and the kinetic energy stays uniformly bounded, but the maximum velocity grows without limit as the time approaches one. The same construction can be adapted to every positive viscosity.

The inclusion of external forcing is therefore not an improvised loophole. It is explicitly permitted by the problem Clay published. It does narrow the physical interpretation, however. OpenAI has not shown that every turbulent flow breaks down, or that an ordinary unforced fluid will necessarily form a singularity.

The Claimed Proof Constructs a Finite-Time Singularity

The central object in the proof is an axisymmetric vortex. Fluid spirals inward toward a vertical axis while flowing outward in opposite axial directions above and below a dividing layer. Carrying angular momentum toward a smaller radius spins the fluid faster, while axial outflow prevents material from accumulating around the center.

As the singular time approaches, both the radius and height of the intense-flow region decrease. The radius contracts more quickly, so the core becomes proportionally thinner and more column-like. OpenAI compares the geometry to an increasingly elongated strand of spaghetti, although the entire core is shrinking rather than growing longer.

Velocity inside this diminishing region increases without bound. That does not require infinite total energy because the fast-moving fluid occupies an ever-smaller volume. In the paper’s scaling, the kinetic energy within the core actually tends toward zero even as the highest velocities diverge.

Viscosity ordinarily smooths velocity differences, which is one reason the problem has resisted proof for so long. OpenAI’s construction balances inward transport against viscous diffusion. The angular Reynolds number increases without bound, meaning that the fluid completes more rotations during the time viscosity needs to diffuse motion across the core.

Building a rapidly spinning vortex is only part of the argument. For any chosen flow, a mathematician could define an external force as whatever residual makes the Navier–Stokes equation hold. Such a force would usually become singular along with the velocity, failing Clay’s conditions.

The paper’s harder step is to keep the external force smooth. It introduces spatially oscillating pulses around the vortex. Their nonlinear momentum flux cancels the singular part of the residual left by the main flow. Further corrections remove smaller errors, after which the construction is localized to produce a smooth, compactly supported force.

Ten Thousand AI Agents Coordinated for 88 Hours

OpenAI began training the unnamed internal model on August 28. On September 1, after hearing rumors that researchers had resolved two Millennium Prize Problems, the company started evaluating the model across the remaining problems and several related questions.

Different agent groups received different variants of each problem. For Navier–Stokes, some pursued proofs of global smoothness while others searched for counterexamples satisfying alternatives C and D. The agents could access a cached copy of the internet, execute code and communicate with other agents in their assigned groups.

A smaller group of nearly 100 agents first spent roughly 50 hours constructing a finite-time blowup result for the inviscid Euler equations. OpenAI then shifted more resources toward Navier–Stokes, supplied the Euler result to the new groups and upgraded the agents when a later training checkpoint became available.

Codex consolidated promising ideas from different groups and fed them into subsequent rounds. The successful Navier–Stokes effort involved around 10,000 concurrent agents and finished on September 5, about 88 hours after the first agents began. For the Navier–Stokes work alone, OpenAI reports 2.7 million inter-agent messages and approximately 130 billion output tokens.

This was not a chatbot producing a proof in one response. It was a large research system combining a frontier model, massive parallel search, tool use, communication between groups, human resource allocation and machine proof checking.

OpenAI says it maintained its standard frontier-evaluation safeguards, including monitoring and isolation, throughout the project. The announcement does not provide enough operational detail to evaluate those controls independently. It also gives no architecture, public benchmark table or release schedule for the next-generation model, beyond claiming a step-function performance improvement and greater capability than GPT-6 Astra.

Lean Makes the Result Auditable, Not Automatically Accepted

OpenAI’s public Navier–Stokes and Euler repository contains a Lean 4 formalization of the main breakdown results. Its project metadata describes the work as a full formalization with no sorry placeholders, using Mathlib and only Lean’s standard logical axioms. It also provides Comparator challenges intended to check the final theorem against independently written versions of the Millennium Prize statements.

That substantially strengthens the claim. If the repository builds successfully, Lean’s kernel can verify that the formal conclusion follows from the definitions and earlier steps encoded in the project. Reviewers do not have to trust millions of lines of informal model reasoning or manually inspect every algebraic manipulation.

Formal verification does not eliminate every question. Mathematicians still need to confirm that the formal definitions accurately represent the analytical assumptions, that the machine theorem corresponds to Clay’s C and D alternatives, and that the informal paper explains the construction faithfully. Independent groups must also reproduce the build rather than relying on OpenAI’s report.

Clay does not accept proposed solutions directly. Under its Millennium Prize rules, a solution must appear in a qualifying outlet, remain published for at least two years and achieve broad acceptance before the institute will consider it. OpenAI says it does not plan to claim the $1 million prize.

A Credit Dispute Complicates the Announcement

OpenAI acknowledges that it began the evaluation after hearing rumors of breakthroughs by NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge. Their related work extended a program developed by Diego Córdoba and Luis Martínez-Zoroa for constructing singularities in fluid equations.

In a public statement about the dispute, Buckmaster said he and Alpöge had used Claude, Codex and other AI tools while working on smooth-forced blowup results for the Euler, Boussinesq and incompressible porous-media equations. He said their drafts and arguments had passed through Codex sessions and questioned whether private research data could have influenced OpenAI’s effort.

Buckmaster also alleged that OpenAI researchers proposed coordinated publication and authorship arrangements, and that an intimidating remark was made after he refused. He carefully stopped short of accusing the company of taking the research, writing that he did not know whether their data had been used.

OpenAI says neither its researchers nor its agents accessed the pair’s specific work before its public release. It says its Navier–Stokes construction and its unforced Euler proof differ from Buckmaster and Alpöge’s forced Euler result. At the same time, OpenAI says it cannot completely rule out the possibility that de-identified data derived from their product usage contributed to model training.

The available evidence does not establish that OpenAI copied their work. It does reveal a serious governance problem for AI-assisted science: researchers may depend on tools operated by organizations that possess far greater computing resources and can become direct competitors after learning that a promising line of research exists.

The Immediate Impact Is Mathematical, Not Industrial

The claimed proof is not a general formula for solving arbitrary fluid flows. It does not provide a faster computational fluid-dynamics simulator, eliminate numerical weather models or show that normal fluids routinely reach infinite velocity.

Instead, it constructs a carefully engineered example demonstrating that the mathematical continuum model can fail under smooth forcing. A physical fluid could not literally attain infinite speed. The singularity marks the point where the idealized equations cease to provide a physically meaningful description and a more microscopic model would be required.

The larger practical result may concern AI research itself. An internal model, orchestrated across thousands of agents, appears to have generated a long analytical construction and translated it into a machine-checkable proof in less than a week. The compute requirements were immense, but the elapsed time was tiny compared with the decades mathematicians have spent on Navier–Stokes regularity.

Final Thoughts

As of September 9, the responsible description is that OpenAI has published a serious, unusually auditable claimed solution, not that the mathematics community has already declared the Navier–Stokes problem solved. The paper, formal theorem and source repository give experts far more to evaluate than a benchmark score or private demonstration.

If the proof holds, the most consequential part may be the research process that produced it. OpenAI did not merely ask a stronger model a difficult question. It used 10,000 agents, 130 billion output tokens, iterative model checkpoints, cross-group synthesis and formal verification to turn a rumor about a promising research direction into a candidate Millennium Prize resolution within 88 hours.

That capability also makes the credit and data dispute impossible to treat as a side issue. AI labs are becoming scientific collaborators, infrastructure providers and well-funded research competitors at the same time. The decisive next step is independent mathematicians reproducing the Lean build, reading the analytical argument and determining whether it correctly settles the problem Clay posed.