A Weekend Rumor Started a Multimillion-Dollar Sprint
OpenAI's race to solve a piece of the Navier-Stokes Millennium Prize Problem did not start with a research plan. It started with a rumor.
In late August 2026, word reached OpenAI that a mathematician working with an Anthropic researcher was closing in on a solution to a related open problem, the forced Euler equations. Diego Cordoba and Luis Martinez-Zoroa had built the underlying mathematical approach; Tristan Buckmaster of NYU's Courant Institute and Anthropic's Levent Alpoge had been extending it since mid-August, making real progress.
OpenAI's response was to scale up fast. Researchers first ran roughly 1,000 agents for 50 hours against the Euler problem, then expanded to about 10,000 concurrent agents for an 11-hour push on the harder Navier-Stokes case. OpenAI Chief Research Officer Mark Chen later put the total compute cost in the range of millions of dollars. The company reached an answer on September 5 and announced it on September 8, roughly 88 hours after the effort began.
What OpenAI Actually Claims to Have Solved
The result OpenAI announced is narrower than it sounds. It addresses one specific variant, blow-up under smooth forcing, produced by an internal, unreleased model OpenAI says is more capable than its public GPT-6 Astra line, running as a coordinated swarm of agents rather than a single pass.
OpenAI paired the announcement with a machine-verified formalization of the proof in the Lean theorem prover, evidence the logical steps hold together mechanically. That is not the same as peer review. The Clay Mathematics Institute, which oversees the seven Millennium Prize Problems and the one million dollar award attached to each, has not verified or accepted the result, and independent mathematicians have not finished reviewing the human-readable argument behind the Lean proof.
| Phase | Agents | Duration | Target |
|---|---|---|---|
| Phase 1 | about 1,000 | 50 hours | Related Euler equations |
| Phase 2 | about 10,000 | 11 hours | Navier-Stokes blow-up case |
| Total | up to 10,000 | 88 hours | Announced September 8 |
The Mathematician Who Says His Drafts Were Read
Tristan Buckmaster's account of what happened next is the part of this story OpenAI has not disputed in public. He says he asked OpenAI researcher Sebastien Bubeck directly whether OpenAI's model had been trained on, or had access to, the private Codex sessions where he and Alpoge had been storing their own drafts on the related problem. Bubeck denied it the first time Buckmaster asked. The second time Buckmaster asked, specifically about training data, he says he got no answer.
Buckmaster also says Bubeck offered him a choice: publish first with OpenAI credited alongside him, or keep sole claim to recognition on the result, but only if he removed Alpoge's name, because of Alpoge's Anthropic affiliation. When Buckmaster refused to drop his collaborator, he says Bubeck told him, "Why would you ruin your career?" and later, "If you don't want me to be nice, then I don't have to be nice."
Neither OpenAI nor Bubeck has published a point-by-point response to these specific claims.
Twenty-Five Fields Medalists Draw a Line
Three days after OpenAI's announcement, Terence Tao, the 2006 Fields Medalist, published "A Severe Misalignment of AI in Mathematics" on his blog, co-signed by 24 other Fields Medal recipients spanning ceremonies from 1978 to this year, among them Peter Scholze, Maryna Viazovska, Caucher Birkar, June Huh, Martin Hairer, Maxim Kontsevich, James Maynard, Manjul Bhargava, and 2026 medalist Yu Deng.
Their argument is not that AI cannot do mathematics. It is that solving a named problem was always meant to be a proxy for conceptual understanding, not the goal itself, and a race to produce the proxy faster can destroy the slower process, the talks, writeups, attribution, and simplification, that turns a solved problem into shared knowledge mathematicians can build on. Rushed, disputed announcements skip exactly that process.
The declaration warns that mass-producing true or false statements "could destroy fertile ground instead of breathing life into new ideas," and that without mathematicians doing the work of integrating a result into the field's own canon, "the crucial human transmission chain between mathematicians would be lost."
What This Means the Next Time a Vendor Announces a Breakthrough
Every part of this sequence, the rumor, the scramble, the announcement, the dispute, happened faster than anyone outside the two labs could check any of it. That is the pattern to watch, not just in mathematics.
An AI vendor's benchmark claim, whether it is a math proof, a coding score, or a safety evaluation, is a press release until an independent party with no stake in the outcome has reviewed it. Eighty-eight hours produced an announcement. It did not produce a verdict, and the people best placed to give one have just said, in public and by name, that the race to get there is part of the problem.
Read next: The Best AI Hacking Tools Are Now Invitation-Only | Three AI Vendors Went Down in One Window on Sept 3
