OpenAI presents a Navier–Stokes proof generated by roughly 10,000 agents

The most important claim an AI laboratory made this week is neither a new score nor a product feature. OpenAI has published a 166-page proof that purports to resolve the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems, along with a Lean certificate. The system that found the construction uses an internal model the company describes as “significantly more capable” than GPT‑6 Astra, and it is not publicly available.

If the argument holds, it would be a historic mathematical result and a much stronger signal of AI capability than completing an exam or improving a bound selected by the provider itself. But three different states coexist today: OpenAI presents a proof, the formal code makes it possible to begin checking that proof, and the Clay Mathematics Institute still lists the problem as “Unsolved”. Saying AI “solved it” without that sequence compresses precisely the part that now matters.

The exact shape of the claim

The Navier–Stokes equations describe the motion of a continuous fluid with viscosity. In three dimensions, it was not known whether perfectly smooth data and forces always produce a smooth solution for all time, or whether velocity can become unbounded at a finite instant. The new construction begins with a fluid at rest, applies a force that is smooth and compactly supported in space and time, and makes velocity diverge as time approaches one while total kinetic energy remains bounded.

The external force is not a loophole added after the fact. Charles Fefferman’s official formulation allows the problem to be resolved either by proving global existence without forcing—options A or B—or by constructing a smooth force for which no global smooth bounded-energy solution exists—C or D. OpenAI’s main theorem claims C in the whole space and, through compact support, D on the periodic torus.

That precision also marks what has not been shown. This is not a general formula for predicting every turbulent flow, nor proof that ordinary water reaches infinite speed by itself. It is a carefully constructed counterexample for the ideal equations, with an equally engineered force. It does not immediately change a weather simulator or a wing calculation; if correct, it establishes that global smoothness is not guaranteed under one of the alternatives expressly accepted by the problem.

The paper’s physical idea is a vortex whose core contracts faster in radius than in height. As it concentrates, azimuthal and axial velocities rise, even as the active region becomes so small that total energy stays bounded. The obstacle is that one cannot simply define a flow that blows up and call the equation’s residual the “force”: that residual must remain smooth too. The construction adds oscillatory pulses whose momentum transport cancels the singular parts, then chains corrections until the force stays smooth through the blowup time.

This is an analytical proof, not a numerical simulation that gets very close to a singularity. That difference explains both its importance and the difficulty of reviewing it: the cancellations, supports, and limits must work across infinitely many scales, not only at a computer’s resolution.

Ten thousand agents are not ten thousand reviewers

The discovery account describes a research machine on another scale. OpenAI began on September 1 by distributing several Millennium Problems and related questions among groups of agents. Nearly one hundred agents first found an unforced Euler blowup in about fifty hours. The team transferred that result to Navier–Stokes, concentrated resources there, upgraded the agents when a more trained version of the model became available, and used Codex to consolidate findings across groups.

According to the company, the decisive group grew to roughly 10,000 concurrent agents. The construction emerged after 88 hours; GPT‑6 Astra took another 17 to formalize it. The Navier–Stokes effort alone produced 2.7 million messages and about 130 billion output tokens. The full set of attempted problems consumed 300 billion.

Those figures are not a benchmark, and they do not measure ten thousand independent ideas. They describe search budget: many branches can repeat errors, follow dead ends, or share the same model and orchestration bias. We also do not know the complete cost, the failed results by problem, how much people guided each reallocation, or a general evaluation of the model that would show whether the capability transfers beyond this project. “More capable than Astra” is, for now, a provider comparison—not a system another laboratory can test.

There is a demonstrable practical change in the type of artifact delivered. The public repository aligns its main theorems with an independent formalization of Clay’s statement, declares zero sorry placeholders in the solutions, and limits axiom dependencies to standard Lean extensions. It also includes Comparator configurations that export the certificate to another checker. That radically narrows the space for an omitted logical step.

It does not eliminate mathematical review. The repository’s own metadata labels its review status “self-assessed.” A kernel can confirm that a term proves an exact formal statement under a set of definitions; specialists must still check that those definitions match the ordinary problem, that the analytical bridges express the right hypotheses, that relevant literature has not been missed, and that the readable argument genuinely explains the certificate. Lean turns a huge proof into an object that can be attacked precisely; it does not turn its generator into an independent reviewer of itself.

Clay’s rules make that institutional distance visible. Before it will consider a solution, the work must appear in a qualifying outlet, at least two years must pass, and it must receive general acceptance from the global mathematics community. OpenAI says it will not claim the prize. As of September 10, the manuscript and repository have been public for two days: there is exceptional evidence with which to begin an audit, but nowhere near enough time to finish one.

Provenance is part of the system too

The proof did not emerge from asking an isolated model to start from nothing. The paper itself acknowledges the line opened by Diego Córdoba and Luis Martínez-Zoroa, whose 2023 work constructed three-dimensional Euler singularities with a less regular force, and the later extension to hypodissipative Navier–Stokes. OpenAI retains the cross-scale amplification of that program, though its oscillatory pulses play a new role: producing the mean flux that cancels the singular force around the vortex.

There was concurrent work as well. Tristan Buckmaster and Levent Alpöge were using several models to carry the Córdoba–Martínez-Zoroa program to smooth forcing in related problems; they released results for Euler, Boussinesq, and porous media shortly before OpenAI’s announcement. In his public statement, Buckmaster says he does not know whether their data were used and explicitly makes no accusation, but disputes the sequence and conversations that precipitated both releases. OpenAI confirms that rumors of that advance motivated its campaign, says it did not access specific user data, and acknowledges that it cannot rule out an indirect influence from de-identified data used to improve its models.

There is no public evidence with which to settle that dispute today. There is enough to draw a less comfortable lesson: in AI-assisted research, the system’s capability includes who identifies a fertile question, which prior work is incorporated, what usage data may feed the model, who directs the compute, and how credit is assigned. Provenance is not a social footnote attached to a proof. It determines whether other people can trust a tool with unpublished ideas when the maker of that tool also competes to produce results.

My reading is that, even with all those reservations, the capability jump is difficult to minimize. An unreleased model, an enormous fleet of agents, and subsequent formalization have produced in days an object that requires specialists to review a possible first-order solution. But the achievement does not show that merely multiplying agents is enough. It shows an industrial chain: human direction chooses and redirects problems, a shared model searches at scale, prior mathematics narrows the space, and a formal checker raises the floor of correctness.

If the proof survives review, the bottleneck will have shifted abruptly. It will no longer be only finding a proof, but absorbing millions of generated steps, reconstructing their ideas, checking formal fidelity, recognizing antecedents, and deciding what deserves to become shared knowledge. The right response is neither to discount the result because a machine produced it nor to declare the case closed because Lean accepted a file. It is to build independent review capacity and provenance rules commensurate with the generation capacity this episode has just displayed.

Sources

• OpenAI, On the Navier–Stokes Millennium Prize Problem, September 8, 2026.
• OpenAI, Finite Time Blowup for Navier–Stokes, technical manuscript, September 8, 2026; and Lean formalization.
• Charles L. Fefferman, Clay Mathematics Institute, Existence and Smoothness of the Navier–Stokes Equation; and Millennium Prize Problem rules, accessed September 10, 2026.
• Diego Córdoba and Luis Martínez-Zoroa, Blow-up for the incompressible 3D-Euler equations with uniform force, 2023.
• Tristan Buckmaster, statement on the joint work with Levent Alpöge and the announcement timeline, September 7, 2026.
• Quanta Magazine, AI Has Solved One of Math’s $1 Million Millennium Prize Problems, September 8, 2026.