Clay says OpenAI's Navier-Stokes result has apparently settled a Millennium problem

OpenAI says its internal research system produced an analytical construction in which an initially smooth three-dimensional incompressible fluid, starting from rest and subject to a smooth external force, develops a singularity in finite time while its total energy remains bounded. The company released both a conventional mathematical write-up and a Lean formalisation of the claimed proof.

That distinction about the external force matters. Clay’s official formulation gives solvers four routes, A through D. A and B ask for global smoothness with zero forcing, while C and D ask for a smooth initial field and smooth forcing for which global smooth solutions break down. OpenAI says its construction establishes C and D. It therefore addresses the Millennium Prize formulation without showing that an unforced fluid must develop a singularity.

OpenAI says it launched the effort on 1 September after an internal model showed unusually strong mathematical performance. The Navier-Stokes search eventually involved on the order of 10,000 concurrent agents, with groups exploring different proof directions and Codex consolidating useful intermediate results across groups. OpenAI reports that the effort used about 2.7 million agent messages and roughly 130 billion output tokens.

According to OpenAI, the agents reached the mathematical construction after about 88 hours. Formalisation and verification in Lean took another 17 hours using GPT-6 Astra. OpenAI also says the internal model driving the search is significantly more capable than Astra and was still being trained when the work was published. Those compute and capability descriptions are OpenAI’s own measurements and have not been independently audited.

The strongest external signal comes from the Clay Mathematics Institute itself. On 11 September, Clay said the Navier-Stokes problem had “apparently been settled” and described the mathematical community as beginning to analyse and interrogate the work. It also stressed that evaluating the achievement and assigning credit would be deliberately unhurried. No Millennium Prize has been awarded, and OpenAI says it does not intend to claim the prize. The proof was released by OpenAI as a manuscript with a Lean formalisation; it is not a peer-reviewed journal publication, and Clay's formal prize evaluation remains pending.

The construction also sits inside a fast-moving line of human and AI-assisted work. Quanta reports that it builds on analytical techniques developed by Diego Córdoba and Luis Martínez-Zoroa, while Tristan Buckmaster and Levent Alpöge announced related AI-assisted Euler results around the same period. OpenAI says an internal investigation found Buckmaster’s earlier Codex prompts could not have influenced its model or result. That is OpenAI’s account, and questions of mathematical validity, provenance and historical credit remain distinct.

If the result survives normal mathematical review, it is a major result in partial differential equations and a striking demonstration of AI-assisted theorem discovery. OpenAI’s system did more than retrieve or formalise an existing proof: the company says a large population of agents searched for a construction, combined intermediate ideas and then passed the result through formal verification.

The theorem does not imply an immediate change to aircraft design, weather forecasting or ordinary fluid simulation. The singular solution is a mathematical construction in an idealised continuum model, and the unforced global-regularity question represented by Clay’s A and B statements remains distinct. The immediate significance is mathematical and methodological: a frontier AI system appears to have crossed from assisting with difficult proofs into producing a result that the institution responsible for a Millennium Prize now describes as an apparent settlement.