OpenAI Solved the Wrong Navier-Stokes Problem

There are two stories here and the louder one is the wrong one.

On September 8, OpenAI announced that agents running on an unreleased internal model had produced a proof that the Navier-Stokes equations can develop a singularity in finite time. (cite index="3-4">OpenAI deployed around 10,000 agents on August 28 and solved it within 88 hours. (cite index="4-10">The resulting proof runs 165 pages. (cite index="20-2">It has been formally checked in Lean, giving mathematicians some confidence it is correct. The press called it a triumph. The internet called it a controversy. Both missed the structural problem underneath.

The credit dispute is real and ugly. (cite index="10-4">NYU mathematician Tristan Buckmaster, who won a 2019 Clay Research Award for proving Navier-Stokes solutions aren't unique, published a statement on September 8 accusing OpenAI of pressuring him over credit for a related, unpublished result. (cite index="9-16">Buckmaster reports pressure over publication and authorship, including pressure to exclude Alpöge because of his Anthropic employment, and comments he understood as career threats. (cite index="8-4">OpenAI's Sébastien Bubeck responded publicly calling the allegations "false and inflammatory." (cite index="29-9">OpenAI acknowledged it intensified work on the Navier-Stokes problem on September 1, after learning other mathematicians were pursuing a similar direction.

Watch the dispute. Do not let it eat the actual story.

The Proof Does Not Claim What the Prize Requires

Here is what nobody is saying loudly enough: (cite index="24-10,24-11">the Clay Mathematics Institute defines the Millennium Prize criteria based on the unforced Navier-Stokes equations, whereas the OpenAI result specifically addresses the forced version. Because the forced result does not satisfy the criteria, the prize remains unclaimed.

Translation: OpenAI ran 10,000 agents for 88 hours, generated a 165-page Lean-verified proof, and did not technically solve the problem with the million dollars attached.

This is not a small caveat. It is the whole scoreboard.

(cite index="27-1,27-2,27-3">What Lean verification guarantees is syntactic and deductive correctness — whether the stated theorem is logically derived from the defined axioms. Lean cannot judge whether the formalized statement matches the physical and mathematical requirements as specified by Fefferman. A slight discrepancy in decay conditions, the definition of forcing smoothness, or the function space means the Lean code could compile successfully yet still not constitute a true solution to the Millennium Problem.

(cite index="27-4">Fields Medal laureate Terence Tao of UCLA praised Buckmaster and Alpöge's result as "a breakthrough across three simplified models — IPM, 2D Boussinesq, and 3D Euler" — while pointing out they themselves had not reached blowup for the Navier-Stokes equations proper. (cite index="26-1,26-3">Tao lamented on social media about AI companies using these longstanding mathematical challenges as marketing proof points, noting AI was being used autonomously to produce answers without providing much insight.

The Model Is Proprietary. So Is the Claim.

(cite index="6-16">OpenAI used an internal model that is significantly more capable than GPT-6 Astra. That model is not public. (cite index="18-1,18-2">The proof reportedly used around 10,000 AI agents over 88 hours, followed by 17 hours of Lean formalization. The claim now awaits independent mathematical verification and potential review by the Clay Mathematics Institute.

In mathematics, an unverifiable claim by an unreleased system is not a proof. It is a press release with Lean attached.

(cite index="28-7">A proposed proof, a verified mathematical result, and an accepted solution to a Millennium Prize Problem are different stages of the same story. OpenAI conflated all three in its announcement, and most coverage followed.

The most likely explanation is not that OpenAI cheated, and not that Buckmaster is wrong to be angry. It is that frontier AI incentives and academic mathematics norms are now on a direct collision course, and no institution has written the rules for impact. (cite index="30-6">Tao argued that AI labs deploying massive, on-demand compute the instant they hear a rumor changes incentives for how openly researchers share promising work — a dynamic that extends well past mathematics into any domain where being first now matters more.

Take that seriously if your research strategy or product roadmap assumes hard problems stay hard for predictable periods. (cite index="8-10">Alpöge and Buckmaster made the case that their collaboration — and its underlying motivation of keeping major math results in human, academic hands rather than corporate ones — predates this week's dispute by a full year. The compute showed up anyway.

What to Watch

Watch whether the Clay Mathematics Institute demands the forced-versus-unforced distinction be resolved before engaging with OpenAI's claim at all. Watch whether OpenAI releases the underlying model or hides behind competitive sensitivity. Watch whether Buckmaster's allegations become a legal dispute over AI-assisted research credit; (cite index="28-8">the question of how discovery should work when tools researchers depend on belong to companies pursuing discoveries of their own needs an answer, and courts are less patient than math journals. Watch the other five Millennium Problems. Every lab now has a roadmap and a motive.