OpenAI’s Millennium Prize fight

The Clay Mathematics Institute still has a million dollars on the table for Navier–Stokes existence and smoothness — one of seven Millennium Prize problems. This week NYU's Tristan Buckmaster and Anthropic mathematician Levent Alpöge posted AI-assisted proofs that edge toward it. Hours later, OpenAI said an unreleased next-generation model had finished the full problem after a week of agentic search. The equations may be the story of the decade. The fight over who got there first — and how — is the story of the week.

What was actually claimed

Buckmaster and Alpöge announced three results on cousins of Navier–Stokes — including 3D Euler and Boussinesq under smooth forcing — with Lean formalization for parts of the work and heavy use of both Codex and Claude. Fields Medalist Terence Tao called the program a remarkable achievement. Separately, OpenAI published what it describes as a complete proof of the forced Navier–Stokes Clay problem, produced by a research model that spent roughly 300 billion output tokens — about $22.5 million at current Astra rates if you bill the run like a product. Clay has not awarded anything. Peer review and independent formalization will decide whether either claim sticks.

The priority fight

Buckmaster says rumors of their approach reached OpenAI before the papers went public, that OpenAI's sprint started only after that tip, and that the lab took the same uncommon route through a smooth external force — options C and D in Fefferman's statement of the problem — that almost nobody else was quietly pursuing. OpenAI's own write-up confirms the latest push began around September 1, inspired by rumors that a Millennium problem had been cracked, and that researchers spoke with Buckmaster and Alpöge. OpenAI also says its team and agents did not see the pair's unpublished work, while admitting it cannot fully rule out that de-identified product-usage data helped improve models. Buckmaster alleges pressure to drop Alpöge from authorship because he works at a rival lab, plus career threats when he wanted the dispute public. OpenAI researcher Sébastien Bubeck has called those allegations false and inflammatory and said a fuller response is coming.

Why this is bigger than one prize

Frontier labs already sell reasoning as a product. This episode is that pitch colliding with academic norms: priority, credit, independence, and who owns a proof when the compute bill looks like a Series A. It also sits next to the same week's other trust stories — a class action over Anthropic's Max plan for power users, and Mistral's €3B Series D framed as European sovereign AI — without needing them to share a plot. Capability is racing. Governance of credit, disclosure, and customer promises is not.

The point

Treat OpenAI's Navier–Stokes post as a disputed claim, not a prize-winning proof. The questions that matter are practical: when did the work start, how much did humans guide the agents, can outsiders check the Lean proof files, and does OpenAI ever use data from customers working on the same problem with its tools? For OpenAI's other trust issue this season, agents that escaped their sandbox, see OpenAI's rogue agents, explained. Related: Language Models, Machine Learning, Ethics.