OpenAI says an unreleased internal model produced a 166-page construction showing finite-time blowup in the three-dimensional Navier–Stokes equations, together with a Lean 4 formalization. If the argument survives independent scrutiny, it would satisfy the counterexample alternatives in the Clay Mathematics Institute’s Millennium Prize formulation.
What changed
The paper constructs smooth, compactly supported forcing for a fluid that starts at rest, keeps bounded kinetic energy, and develops unbounded velocity in finite time. OpenAI also released a public repository containing formalizations for both whole-space and periodic versions of the claimed result.
The company says the work came from an internal model “significantly more capable” than GPT-6 Astra. It reports coordinating roughly 10,000 agents for the Navier–Stokes effort, which ran for about 88 hours, exchanged 2.7 million messages, and generated approximately 130 billion output tokens. OpenAI says GPT-6 Astra then spent another 17 hours on Lean formalization and verification.
The practical consequence
The release provides an unusually concrete artifact for evaluating AI-assisted mathematical research: readers can inspect the analytical argument and the formal certificate instead of relying only on a benchmark score. If validated, it would also show that a very large multi-agent system can help produce a result at the frontier of mathematics. That does not make the method available to ordinary researchers today; the generating model is unreleased, and OpenAI published no price or resource breakdown beyond message and token totals.
Why this is not yet a settled result
A Lean certificate can check a formal statement against definitions and a trusted software stack, but it does not by itself establish that every formal definition captures the intended mathematics or replace expert review of the analytical construction. DrComps has not audited the 166-page proof, built the repository, or obtained an independent mathematical assessment.
The Clay Mathematics Institute still lists Navier–Stokes among its unsolved problems. Its rules require a proposed solution to appear in a qualifying outlet, survive at least two years of rigorous examination, and gain general acceptance in the global mathematics community before the institute considers it. OpenAI says it does not intend to claim the prize.
Benchmark status and limits
OpenAI calls the internal model’s benchmark performance unprecedented but provides no benchmark names, scores, test set, model card, or access for independent reproduction. The 10,000-agent scale and 130-billion-token total are vendor-reported process measurements, not evidence of efficiency or a general-purpose success rate. The construction also uses smooth external forcing; it should not be paraphrased as proof that every ordinary unforced fluid becomes singular.