OpenAI's model proved Navier–Stokes blows up — and Lean checked it

Disclosure: this newsroom's assistant is built by Anthropic, one of whose employees appears in the credit dispute described below.

OpenAI published a proof on 8 September 2026, produced by an internal model, addressing the Navier–Stokes existence-and-smoothness problem — one of the seven Clay Millennium Prize Problems, open for roughly 90 years. The answer is negative: the model constructs a finite-time blowup, a configuration in which a vortex tightens and spins ever faster while the fluid's total energy remains bounded. OpenAI says the run took 88 hours across as many as 10,000 agents working roughly concurrently, and that the argument was verified in Lean on 6 September. A credit dispute followed: OpenAI says it began on 1 September after hearing a rumour it later connected to Levent Alpöge, an Anthropic employee, and Tristan Buckmaster of NYU; on offering a joint announcement it learned their result concerned the forced Euler equations — related, but a different problem. The tracker records the proof.

Why it matters

Analysis — interpretation, not additional fact.

The Lean verification is what separates this from every previous "AI solves famous problem" cycle, and it deserves to be stated precisely. A Lean-checked proof is machine-verified against a formal kernel: if the statement being proved is the right statement, the proof is correct, and no human referee's fatigue or generosity enters into it. That is a stronger guarantee than most published mathematics carries on the day it appears. It removes the question this newsroom usually has to ask — is the claim real — and leaves a narrower one: is the formalised statement the theorem everyone means by Navier–Stokes.

What it does not do is award the prize. Clay's rules want peer-reviewed publication and a waiting period, and those exist for reasons Lean does not address: whether the formalisation faithfully encodes the problem, and whether the community can read and use the argument. A machine proof that no mathematician has internalised is a settled question and an unabsorbed one. Expect the interesting work of the next year to be humans reconstructing why the blowup happens, not checking whether it does.

The credit dispute is the part to hold lightly and not dismiss. Two researchers were reportedly circulating a related result; OpenAI heard about it, spent five days, and published first on the adjacent-but-harder problem. Nothing there is misconduct on the facts as stated — the problems turned out to be different — but the sequence describes a research environment where hearing that something is possible is worth days rather than years, and where a lab with 10,000 agents can convert a rumour into a formal proof before the originators finish theirs. Whatever that is, it changes what "priority" means in mathematics, and the discipline has not agreed on the rules yet.

What to watch

Whether mathematicians endorse the formalised statement as the Millennium problem, whether Clay engages with a machine-produced proof at all, and whether Alpöge and Buckmaster's Euler result publishes independently.