OpenAI maintains that one of its models solved the Navier-Stokes problem, one of the seven Millennium Prize Problems, each carrying a one-million-dollar prize. The company says it took 88 hours of compute, according to reports published on September 11 and 12, 2026.
Behind the announcement lies a dispute over credit. Tristan Buckmaster, a professor at New York University, told The Verge that OpenAI offered him practically unlimited compute and sole authorship of the paper if he left out Levent Alpöge, a researcher at Anthropic and his collaborator on the same problem. A company spokesperson, Laurance Fauconnet, denied that the messages Buckmaster wrote in Codex over the previous two months influenced the system, including its training. Buckmaster responded that statements of that kind should be taken with skepticism. The seven problems were posed in 2000, and only one, the Poincaré conjecture, has been solved since.
Against that backdrop, TechCrunch reported that 25 Fields Medal-winning mathematicians signed an open letter this week warning that artificial intelligence labs threaten their intellectual work. The signatories wrote that these solutions are often announced in haste, with no time for careful drafting, the isolation of new methods or the citation of prior work, which raises severe questions of attribution and plagiarism. OpenAI's proof still lacks independent verification. The company said it used nearly 10,000 agents and tens of millions of dollars in compute. Sébastien Bubeck, a researcher at OpenAI, rejected Buckmaster's account and told the New York Times that Alpöge's affiliation with Anthropic was a point of friction. OpenAI also withdrew its sponsorship of a mathematics event at Caltech this week.
The result now rests with the mathematical community: without independent verification and without full disclosure of the method, it does not enter the canon. The letter includes recommendations for mathematicians, institutions and legislators, and follows the Leiden Declaration of June 2026. How a finding is credited when a machine produces it remains an open question.
This article was written with the assistance of artificial intelligence from verified sources and reviewed by a human editor before publication.