OpenAI Claims an AI System Found a Navier–Stokes Blow-Up Proof

OpenAI has made an extraordinary mathematical claim: an internal, unreleased AI system using roughly 10,000 concurrent agents found a proof that certain three-dimensional incompressible Navier–Stokes flows develop a singularity in finite time. OpenAI says the agent run took about 88 hours, and a separate formalization and verification in the Lean proof assistant took another 17 hours. 17

That is not the same as an established resolution. OpenAI has released a written proof and Lean material for scrutiny, but outside mathematical acceptance—and the Clay Mathematics Institute’s prize process—take time. 126

What OpenAI says it proved

The Navier–Stokes equations describe the motion of viscous fluids through a velocity field, pressure, viscosity, nonlinear transport, and—in the Clay formulation—an allowed forcing term. The famous three-dimensional regularity question asks whether smooth, physically admissible flows must remain smooth for all time or whether smooth data can lead to a finite-time breakdown. 172133

OpenAI’s stated result is a blow-up construction: a solution beginning from rest, driven by a smooth, compactly supported external force, whose velocity becomes unbounded at a finite time while kinetic energy remains bounded. 134

In ordinary language, the claimed example says that viscosity does not always prevent a three-dimensional fluid flow from becoming mathematically singular. The result is not simply a numerical simulation of extreme turbulence; it is presented as an analytical proof, with its logical steps translated into Lean for machine checking. 134

Why a valid proof could be historic

Navier–Stokes existence and smoothness is one of the seven Millennium Prize Problems, for which the Clay Mathematics Institute established a $7 million fund divided into $1 million prizes. 1820

The official problem permits two broad routes: prove global smooth solutions under the required conditions, or construct an admissible finite-time breakdown example. Clay’s official description includes conditions on both initial data and external forcing. 1721

OpenAI says its construction establishes statements C and D of that formulation—the finite-time singularity alternatives with smooth forcing in Euclidean and periodic settings. 17 If that claim withstands expert review and matches the formal problem requirements, it would potentially be only the second Millennium Prize Problem resolved, after the Poincaré conjecture; six were still described as remaining when OpenAI announced the result. 1118

Why Lean verification is important—but not the final word

Lean is a proof assistant: it checks whether a formal statement follows from definitions, axioms, and prior verified results encoded in the system. A successful Lean verification is therefore meaningful evidence that the formalized proof has no logical gap at the level Lean checked. 17

But mathematical validation still requires experts to assess the whole chain:

• whether the formal theorem exactly captures the claimed Navier–Stokes setting;
• whether the translation from the written argument to formal definitions is appropriate;
• whether the assumptions satisfy the Clay problem’s conditions; and
• whether the broader mathematical community accepts the proof.

Those are not mere procedural details. A machine check validates a specified formal object; it does not independently decide whether that object answers every intended interpretation of a prize problem.

Why Clay has not awarded the $1 million prize

Clay does not accept direct submissions of proposed solutions. Before it will consider one, the proposal must be published in a qualifying outlet, at least two years must pass after publication, and the work must receive general acceptance in the global mathematics community. 26

So even if OpenAI’s proof is ultimately correct, an immediate prize award would not fit Clay’s published rules. The appropriate current description is a claimed solution undergoing scrutiny, rather than a prize-winning result. 126

The key scope distinction: forced versus unforced flows

Discussion of the announcement often blurs two different questions. The most familiar popular framing asks whether a smooth, unforced 3D flow can spontaneously lose regularity. OpenAI’s stated construction instead uses a smooth external force. 3435

That distinction matters scientifically, but it does not automatically disqualify the claim from Clay’s framework: the official formulation includes alternatives involving smooth, rapidly decaying forcing. 1735 The real issue for the prize is whether OpenAI’s proof satisfies every condition of the relevant Clay alternative—not whether it matches a narrower informal version of the question.

Why the announcement triggered a priority and data-use dispute

The announcement arrived amid related work by NYU mathematician Tristan Buckmaster and Levent Alpöge, an Anthropic researcher working with Buckmaster in a personal capacity. Their work concerned related forced fluid equations rather than an established solution of the standard Navier–Stokes problem. 237

Buckmaster alleged that OpenAI sought a collaboration arrangement that would exclude Alpöge because of his Anthropic affiliation. He also raised questions about whether the researchers’ interactions with OpenAI products, including Codex, could have contributed to model training. 25052

OpenAI said that its researchers and agents did not access the pair’s specific work or user data in solving the problem. It also said it could not completely rule out that de-identified data derived from product use may have helped improve its models. 5354

Those statements leave important questions unresolved: whether any relevant interactions were eligible for training, whether they were actually used, and whether they could have materially affected the result. Public reporting does not establish that OpenAI used Buckmaster and Alpöge’s work, and the allegation should not be treated as proven. 25354

What to watch next

The decisive test is not the size of the agent swarm or the speed of the result. It is whether independent specialists can inspect the manuscript and Lean formalization, reproduce the formal check, and agree that the theorem meets the relevant Clay conditions.

If that happens, OpenAI’s announcement may mark a landmark in both mathematics and AI-assisted research. Until then, it is best understood as a highly consequential, publicly available candidate proof—alongside a separate unresolved debate about priority, data governance, and how AI companies should handle researchers’ unpublished ideas. 12653