Experts dispute OpenAI’s account of its proposed Navier–Stokes proof

Mathematicians are challenging OpenAI’s explanation of a 166-page proposed solution to the Navier–Stokes problem, even though the proof was written in Lean, a formal language that checks logical steps. The Clay Mathematics Institute is still reviewing the work, so the submission has not been accepted as a solution to the famous open problem.

OpenAI said its argument followed an approach associated with mathematicians Diego Córdoba and Luis Martínez-Zoroa. Several experts disagree. Gonzalo Cao-Labora of EPFL said the detailed method is different and that the resemblance is only superficial. A separate dispute concerns whether the work drew on publicly available experiments by Tristan Buckmaster and Levent Alpöge, which OpenAI denies using.

The project also tested an unusually large multi-agent setup. OpenAI reportedly ran about 10,000 model instances for roughly 88 hours and consumed 130 billion tokens. Researcher Noam Brown estimated that the multi-agent structure accounted for less than 10 percent of the success, but no controlled comparison has established what 10,000 agents added over a smaller group.

Formal verification can establish that steps follow specified rules; it does not automatically make a long proof understandable or historically attributable. Until expert review is complete, the document remains a proposed proof, and claims about both its novelty and its development process remain contested.