NYU Mathematician Says OpenAI Copied His Work

NYU mathematics professor Tristan Buckmaster told CNBC Television that OpenAI's claim to have solved the Navier-Stokes existence and smoothness problem followed directly from work he was developing with collaborator Levent on OpenAI's own Codex system.

The timeline is the entire case.

Buckmaster said that by the end of August he and Levent believed they were roughly a month away from a result on Navier-Stokes, one of the seven Millennium Prize problems for which the Clay Mathematics Institute offers $1 million. At that point a rumor spread from Anthropic to OpenAI that Anthropic researchers were close to solving two Millennium problems, Navier-Stokes and the Hodge conjecture. The rumor jumped to the tech world and then to X, spawned betting markets, and according to Buckmaster triggered an all-out compute push inside OpenAI, which saw a rival solving a Millennium Prize as an existential signal about model quality.

Buckmaster said he was using Codex to interpret and develop work coming from Levent, which meant drafts, prompts and intermediate calculations lived on OpenAI servers. He was on a paid personal account, not an enterprise account, and he discussed the toggle that controls whether OpenAI can train on user data. The default is on. He said he had turned it off on two accounts but had left one account with training on.

After the rumor intensified, he said he and Levent contacted an OpenAI mathematician to calm things, clarifying that they had not solved Navier-Stokes. They had solved the related Euler equation, a step below Navier-Stokes in the blow-up hierarchy, and wanted to avoid being cast as an Anthropic project because Levent works for Anthropic. OpenAI, he said, initially offered tokens and compute to avoid competing, with talk of millions of tokens to finish the project together.

Then came a Saturday warning.

Buckmaster said he was told OpenAI was about to do something stupid and to get on a call. On that call, he alleged, OpenAI offered to let him publish its Navier-Stokes paper as sole author if Levent were removed, which would have positioned him for the prize money, and later pressured him when he called the offer unethical. He quoted an OpenAI staffer identified as Sebastian as saying, "I've been nice to you. I don't have to be nice," and "Why would you ruin your career?" OpenAI has not denied those specific lines, he told CNBC Television. He said the company did acknowledge offering sole authorship.

The technical claim is where the dates tighten. OpenAI first told him it had one-shotted Navier-Stokes, meaning a single prompt containing the prize statement produced the proof. Buckmaster said that claim evaporated within minutes, and internal messages showed work across many problems with agents. OpenAI has since said categorically that Buckmaster's Codex prompts over the last two months could not have influenced the system, including training. Buckmaster noted the window shifted, from April to July in earlier statements to the New York Times, and argued OpenAI would not have needed much. After he revealed his name on Sept. 3, he said, OpenAI went from no progress before knowing his identity to reproducing exactly his and Levent's Euler argument and methods within a day. He described OpenAI using 100 agents to reach their Euler result and 10,000 agents toward Navier-Stokes, landing on the same methods they had left off with.

What had to be true before any of this could happen was mundane. A mathematician using a personal Codex account with default training on, grant proposals and panel summaries routinely pasted into models, and no hard isolation between user data and frontier research. Buckmaster argued that is the real risk, that a provider with visibility into proposals can front-run the next few years of a field. He pointed to OpenAI's dump of hundreds of new mathematics findings this week on algebra, number theory and logic, which he said wiped out research programs for early-career mathematicians on Tuesday. Our earlier look at the proof noted it still required human checking, a point now central to what still has to happen: independent verification, formal review, and a Clay Institute decision. See our prior note OpenAI's Navier-Stokes Proof Still Needs Humans.

Scrutiny has already found cracks. Outside the interview, mathematicians cited by New Scientist say OpenAI mistranslated mathematics into code in publishing its Navier-Stokes proof, a subtle error that undercuts the presentation even as the claimed result awaits peer review.

Buckmaster said he still uses Anthropic and still has Codex installed, but he no longer puts original ideas into it. The question left is not whether a Millennium Prize can be won with agents. It is whether a platform can be both the notebook and the competitor.