OpenAI announced Tuesday that one of its unreleased models resolved a significant aspect of the Navier-Stokes equations — one of mathematics' legendary Millennium Prize problems — using a swarm of roughly 10,000 AI agents working for about 88 hours. The Clay Mathematics Institute, which administers the $1 million bounty, has not yet awarded the prize. OpenAI says it has no intention of claiming it, stating its only goal 'is to report on the substantial progress of our AI models.'
The technical result is real. The Navier-Stokes equations, written down roughly 200 years ago, describe how fluids move. The Millennium problem asks whether those equations always hold or sometimes 'blow up' — producing physically impossible infinities. OpenAI's agents found that blow-ups do occur. According to New Scientist, the specific scenario involves a spinning whirlpool that grows longer, thinner, and more intense over time, with the new proof showing this happens in finite time using only a finite amount of force on the fluid.
The approach built on prior human work. In 2021, Luis Martínez-Zoroa at CUNEF University and Diego Córdoba at the Spanish National Research Council developed a layered method for constructing blow-ups. Earlier this week, Tristan Buckmaster at New York University and Levent Alpöge — a researcher at Anthropic, though not acting on its behalf — published a breakthrough removing a key unrealistic requirement from that framework for Euler equations. Hours later, OpenAI published its full result. 'AI use was absolutely pivotal, and the scale of it is interesting and useful, but the solution approach was developed by brilliant humans,' said Steven Brunton at the University of Washington. John Evans at the University of Colorado Boulder noted that the key ideas the AI agents used are 'relatively simple ones that an undergraduate could have learned in class.'
Here's the thing: the timing and the conduct around the announcement are what have rattled researchers. Buckmaster said he contacted OpenAI to ask when it began working on the problem and what data its model had been trained on. According to Buckmaster, the conversation quickly turned hostile. An OpenAI researcher asked him, 'Why would you ruin your career?' when he said he would go public with what happened. When he pressed further, the reply was: 'If you don't want me to be nice, then I don't have to be nice.' OpenAI also urged Buckmaster to publish the work crediting its internal model and to drop Alpöge as coauthor.
OpenAI has flatly denied using Buckmaster's specific user data, stating: 'We did not see any of their work through any means until they released it publicly.' It did, however, acknowledge it 'cannot rule out that de-identified data derived from their usage of our products helped improve our models.' OpenAI researcher Sébastien Bubeck told a press briefing that the company began working on the problem after finding rumors of competing progress 'on Twitter.' Bubeck, named in Buckmaster's account, has disputed parts of it, denying he ever asked Buckmaster to remove Alpöge as coauthor. Abhishek Saha, a mathematics professor at Queen Mary University of London, said OpenAI engaged in the 'kind of things that mathematicians will generally not do.'
The demo is not the product, and the proof is not the process. OpenAI's model may have compressed months of human-driven progress into days — Zaher Hani at the University of Michigan said exactly that. But a company that spent millions of dollars in a sprint triggered by Twitter rumors, then allegedly threatened a competing researcher, is not behaving like a scientific institution. It is behaving like a platform optimizing for the announcement. The math community deserves better norms than that, and so does the science that depends on it.
