OpenAI claims Navier-Stokes proof; mathematicians accuse it of credit theft

OpenAI announced today that an internally developed model has found a proof for the Navier-Stokes problem, one of the Clay Mathematics Institute's seven Millennium Prize Problems, each carrying a $1 million award. At a press conference, Sebastian Bubeck, a mathematics and AI researcher at the company, said training began on August 28. After rumors surfaced of progress at Anthropic, OpenAI committed massive resources: more than 1,000 agents worked on the problem for over 50 hours, at a compute cost of "millions of dollars," according to Mark Chen, head of research. On Sunday morning, Bubeck said, the final solution arrived, formalized in Lean, a programming language used for formal verification of mathematical proofs.

The competing work surfaces

On Monday, Tristan Buckmaster, a mathematician at NYU, and Levent Alpöge, a researcher at Anthropic, published documents describing significant progress in a related area, the unforced Euler equations. The pair noted they had used several models, including Claude and Codex, to complete the work. In a separate statement, Buckmaster claimed that last week he learned OpenAI was aware of their work and had begun investing significant resources in solving the problem.

The allegations: log access and an offer to omit a name

Buckmaster added that he asked OpenAI leadership whether they had been exposed to the pair's Codex logs. He said he was told the model "did not access user data," but the company did not respond to questions about training. He further claimed OpenAI made several "offers," among them the possibility of publishing a paper announcing the problem's solution by an internal company model, without Alpöge's name appearing. Bubeck denied on social media that the company had proposed removing Alpöge's name, and Sam Altman, OpenAI's chief executive, intervened to defend the team's work.

OpenAI's response: we didn't see the material, but can't rule out indirect influence

In a briefing, Bubeck and other executives denied that the company or its agents had accessed the pair's work before its public release on Monday. However, an official company blog post stated that "while unlikely, we cannot rule out that semi-identified data derived from their use of our products helped improve our models." Bubeck emphasized: "We acknowledge Alpöge and Buckmaster's priority in work on unforced Euler, and have only congratulations on the monumental achievement. We did not use their prompt or proof to guide the models or agents." Ven Chandrasekaran, another mathematician at the company, noted that OpenAI's solution is fundamentally different in nature from the output produced by the pair's model.

What this means in practice

Neither group has responded to a request for comment from WIRED. OpenAI's proof has not yet undergone formal peer review, and the Lean publication allows mechanical verification, but the mathematical community will need to determine whether the solution actually closes the problem in full or addresses only a partial version. For now, the episode illustrates how the capability race among large AI labs is spilling into the arena of scientific credit, with the tools themselves — Codex, Claude — being used by both sides.