OpenAI says a swarm of roughly 10,000 autonomous AI agents produced a proposed solution to the Navier-Stokes existence and smoothness problem, one of seven Millennium Prize Problems set by the Clay Mathematics Institute, each carrying a $1 million award for a verified solution. The company published the proof on September 8, describing it as the output of an internal research system not yet available to the public.
The announcement was almost immediately overshadowed by a dispute over credit. NYU mathematician Tristan Buckmaster said he had made preliminary progress toward the same problem and announced three related proofs of his own on September 8, only for OpenAI’s paper to appear shortly afterward covering similar ground.
What is the disagreement about?
According to Buckmaster’s account, he mentioned his unpublished work to a mathematician affiliated with OpenAI on September 3, amid rumors that Anthropic researchers were also making progress on the same class of problems. Three days later, he says OpenAI researcher Sebastien Bubeck told him in a phone call that an internal OpenAI model had already generated a roughly 100-page proof addressing the same question. Buckmaster has said he was then offered a choice: publish a joint paper a day after his own team’s announcement, or have a write-up credited to him alone with a collaborator’s name left off because of that person’s employer.
OpenAI has not disputed the substance of Buckmaster’s timeline in detail but maintains that its system independently produced a complete proof. The company has not said whether any of its researchers reviewed or drew on Buckmaster’s unpublished work before finalizing its own paper.
The Navier-Stokes equations describe how fluids move and are used across fields from weather forecasting to aircraft design. The Clay Mathematics Institute’s version of the problem asks whether solutions to the equations always remain smooth and well-defined in three dimensions, or whether they can break down. A full proof, if independently verified by the mathematics community, would resolve a question that has stood open since the institute listed it among its seven Millennium Prize Problems in 2000.
Neither OpenAI’s proof nor Buckmaster’s has yet been formally verified by outside mathematicians, a process that for a problem of this complexity can take months or years. The Clay Mathematics Institute has not commented on either submission or indicated a timeline for review.
The episode has added to a broader debate in the mathematics and AI research communities over how credit should be assigned when AI systems and human researchers converge on similar results in a short window, particularly when one side has visibility into the other’s unpublished work.
