The claim from OpenAI is that a system built in house wrote a proof settling Navier-Stokes existence and smoothness, which sits among the seven Millennium Prize Problems the Clay Mathematics Institute set out in 2000. The company’s own account of the result: a fluid that starts out smooth in three dimensions can still lose that smoothness, developing infinite speed at a single point within finite time. That is a negative answer to a question mathematicians had left open for roughly 90 years, not a proof that fluid flow always behaves.
Read carefully, the framing matters. OpenAI is not claiming that smooth solutions persist forever. It is claiming the opposite: under the conditions its system worked out, a swirling vortex can spiral inward, stretch, and speed up without bound, even while a finite amount of energy is doing the work. In the official Millennium Prize language, OpenAI says its system established the statements that correspond to a breakdown, not to permanent smoothness.
The part of this announcement that can actually be checked is the Lean formalization. OpenAI says its system produced both a written analytical proof and a machine-checked version in Lean, a proof assistant that verifies each logical step against a fixed set of rules. A Lean file either compiles against those rules or it fails, which is a different kind of evidence than a company’s prose account of what its model did. Formal verification is currently the only way a claim this large can be tested without waiting years for human referees to work through it line by line, which is exactly why the Lean artifact, not the announcement’s language, is the part worth independent scrutiny.
OpenAI’s account of how the result was produced: a system of coordinating agents, built on a model it says is more capable than GPT-6 Astra, ran roughly 10,000 concurrent agents on the Navier-Stokes problem, sending 2.7 million messages and generating about 130 billion output tokens. The agents reached their result on September 5, about 88 hours after the effort began, and Lean formalization and verification took another 17 hours. OpenAI also says a smaller group of agents, working for about 50 hours, separately produced a disproof of a related but easier question involving the Euler equations, the Navier-Stokes system with viscosity removed.
That Euler side note points to a complication OpenAI discloses itself. It says its effort began after hearing rumors, later traced to Levent Alpöge, an Anthropic employee, and Tristan Buckmaster, a mathematics professor at NYU, that two Millennium problems had already been resolved elsewhere. OpenAI says it reached out once its own Lean verification was done, learned that Alpöge and Buckmaster had solved a different variant (the forced Euler problem), and is crediting them with priority on that specific result. Two independent efforts converging on overlapping problems in the same week, using different methods, is itself notable, but it also means none of this has gone through a single unified review.
What has not happened is the important part to hold onto. The Clay Mathematics Institute runs its own process for recognizing a Millennium Prize solution, involving publication and a waiting period, and none of that has started. Independent mathematicians have not had time to review the proof or the Lean formalization line by line. OpenAI itself says plainly that it does not intend to claim the prize for this result and describes the announcement as a snapshot of its models’ progress rather than a finished, adjudicated achievement.
The comparison worth making is to Anthropic’s announcement, reported here on September 9, that Claude had formalized Andrew Wiles’s existing proof of Fermat’s Last Theorem in Lean. That was translating settled human mathematics into a machine-checkable form. OpenAI’s claim is different in kind: it says its system generated new mathematical content, not just a formal restatement of known work, which is a much higher bar and one that formal verification alone cannot fully clear without a human community checking the underlying reasoning too.
For anyone using this to gauge AI capability rather than mathematical truth, the actionable signal is narrower than the headline suggests. Watch whether independent Lean reviewers can compile and audit OpenAI’s formalization files, not whether the prose proof reads convincingly, before updating any assumption about what these systems can independently discover.
Per OpenAI’s own announcement on openai.com, published this week.
