OpenAI Claims Unreleased Model Solved Navier-Stokes Millennium Prize Problem in 88 Hours, Faces Scooping Allegations

What happened

OpenAI announced Tuesday that an unreleased internal model solved the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize problems carrying a $1 million bounty from the Clay Mathematics Institute. According to OpenAI, the model took 88 hours to produce a solution by coordinating roughly 10,000 AI agents on the task. The problem, concerning the mathematical behavior of fluid flow equations, has resisted human researchers for nearly 90 years.

OpenAI says it does not intend to claim the prize money, and the Clay Mathematics Institute has not yet evaluated or awarded the bounty for this result. The company frames the effort as a demonstration of AI capability rather than a bid for the prize itself.

The timing problem

The announcement came one day after NYU mathematics professor Tristan Buckmaster and Anthropic researcher Levent Alpöge published findings on a related problem. Buckmaster says he contacted OpenAI after learning the company was aware of their progress, asking when OpenAI began working on the problem and what data trained its model. According to Buckmaster, the exchange grew hostile — he says an OpenAI researcher asked, "Why would you ruin your career?" when he indicated he would go public, and later said, "If you don't want me to be nice, then I don't have to be nice." OpenAI reportedly urged him to credit its internal model and drop Alpöge as coauthor.

OpenAI researcher Sébastien Bubeck, named in Buckmaster's account, disputes asking that Alpöge be removed as coauthor. OpenAI has denied accessing Buckmaster's Codex sessions, stating in its blog post: "We (the researchers and the agents) did not see any of their work through any means until they released it publicly — in particular, no specific user data was accessed in order to solve this problem." The company added a caveat: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models," while maintaining the two proofs differ significantly.

OpenAI says it began pursuing Navier-Stokes after seeing rumors on Twitter/X that other researchers were making progress on Millennium Prize problems, and only later realized those rumors referred to Buckmaster and Alpöge's work. By its own account, according to Science's report on a press briefing, the effort cost millions of dollars and was undertaken on short notice once the rumors surfaced.

Why mathematicians are alarmed

Mathematics as a discipline operates on an informal norm of open collaboration. "Mathematics depends heavily on an informal norm of trust," said Matthew Ballard, a mathematics professor at the University of South Carolina and associate director at ICARM. "Researchers routinely share incomplete ideas and ongoing work with colleagues to sharpen their thoughts. It is done with the expectation that it will not turn into a competition."

Abhishek Saha, a mathematics professor at Queen Mary University of London, called OpenAI's conduct the "kind of things that mathematicians will generally not do," noting that scooping is rare partly because cutting-edge problems require specialized expertise few people possess.

Jeremy Avigad, a Carnegie Mellon professor and ICARM director, said even the possibility that AI systems could surface ideas from private queries is "chilling," warning it could make researchers more guarded about discussing unfinished work — a shift he called "sad to think about." Brown University professor Brendan Hassett said OpenAI should be held to a higher standard of proof given "the history of the AI companies appropriating copyrighted work without permission or payment."

What this means

The technical result — if verified by the mathematics community — is a genuine signal that large-scale AI agent systems can now contribute to unsolved problems in pure mathematics, not just competition-style benchmarks. That alone is significant.

But the controversy exposes a governance gap that will outlast this specific dispute. OpenAI's own hedge — that it "cannot rule out" indirect influence from de-identified user data — is an admission that the company cannot fully audit what its models learn from private user sessions, even when it wants to deny wrongdoing. For a research community built on informal trust and open discussion of half-finished ideas, that uncertainty alone changes incentives: mathematicians may now think twice before typing partial proofs into any AI coding or reasoning tool. Whether or not OpenAI accessed Buckmaster's data, the fact that the question cannot be definitively answered is likely to have a lasting chilling effect on how researchers engage with AI products going forward.