NYU mathematician accuses OpenAI of fighting dirty over math proof

NYU mathematics professor Tristan Buckmaster announced three proofs on Tuesday, including a preliminary finding on one of the major unsolved problems in theoretical mathematics, produced in collaboration with Anthropic mathematician Levent Alpöge using both OpenAI’s Codex and Anthropic’s Claude models. The announcement came bundled with a dispute over how OpenAI handled a parallel effort on the same problem.

Buckmaster and Alpöge were finalizing their results when they learned that information about their progress had reached OpenAI. When they asked OpenAI about it, the company said it had already achieved a full proof of the central problem, and grew evasive when asked how much human input was involved and when the work began; Buckmaster says it was eventually agreed that OpenAI’s first prompt on the problem had been sent in the days just before, after information about their work had reached the company. Shortly after Buckmaster’s statement went public, OpenAI published its own full proof of the Navier-Stokes existence and smoothness problem — one of seven Millennium Prize problems, each carrying a $1 million bounty from the Clay Mathematics Institute — crediting an unreleased next-generation model and a week-long effort that consumed 300 billion output tokens, worth $22.5 million at current Astra compute rates. OpenAI’s own post confirmed much of the timeline, saying the latest effort began September 1, prompted by rumors that two Millennium Prize problems had been solved.

Buckmaster said the specific approach he and Alpöge chose — a far less common route through Fefferman’s statement of the problem — was one “almost nobody else” was pursuing, making it suspicious that OpenAI’s team converged on the same tactic at the same time; he said that isn’t a direction a model reaches in days from the bare problem statement alone. Alpöge’s affiliation with Anthropic (he was not working on the project on the company’s behalf) also became a sticking point: Buckmaster alleges OpenAI mathematician Sébastien Bubeck asked him to remove Alpöge’s credit as part of a proposed compromise, and that when Buckmaster pushed to make the dispute public, Bubeck warned him, “If you don’t want me to be nice, then I don’t have to be nice.”

Buckmaster also raised the possibility that OpenAI’s Codex, which he used extensively and which the company can train on unless users opt out, might have let its own models absorb clues from his work. OpenAI’s post said it did not access his team’s specific user data and that its proof differs significantly from theirs, while acknowledging it “cannot rule out” that de-identified usage data helped improve its models.

Why it matters

The dispute suggests that as AI labs race to claim credit for high-profile mathematical results, compute advantage and access to researchers’ own tool usage — not just model capability — increasingly shape who gets there first.