OpenAI says it solved a millennium problem in 88 hours, but the shadow of academic espionage clouds the milestone

By Zendoric · September 10, 2026.

OpenAI announced on Tuesday that one of its internal, as-yet-unreleased models had solved the Navier-Stokes problem: the equations describing the motion of fluids, one of the seven Clay Mathematics Institute "Millennium Prize Problems," carrying a one-million-dollar award and having resisted human researchers for nearly 90 years. According to OpenAI itself, the result came from deploying a swarm of some 10,000 AI agents working in parallel for 88 hours. The company says it will not claim the prize, which in any case the Clay Institute has not yet awarded.

The technical achievement is real, but it arrived wrapped in a controversy that has dominated the conversation more than the result itself. A day before the announcement, New York University mathematician Tristan Buckmaster had published, together with Anthropic researcher Levent Alpöge —acting on his own behalf, not on behalf of his employer— progress on a related problem. Buckmaster recounts that, on learning that OpenAI knew about his work, he contacted the company to ask when it had begun working on the topic and what data it had used to train its model. According to his account, the conversation turned hostile: an OpenAI researcher allegedly asked him, "why would you want to ruin your career?" upon learning that he intended to make the episode public, and when he pressed further allegedly replied, "if you don't want me to be nice, I don't have to be." Buckmaster says OpenAI urged him to publish his work crediting the company's model and removing Alpöge as co-author. Sébastien Bubeck, the OpenAI researcher named in the account, has specifically denied asking for Alpöge to be excluded.

The most sensitive point concerns the data. Buckmaster asked OpenAI whether it had accessed his sessions on Codex —the company's own coding assistant— while he was working on the problem. OpenAI flatly denies any direct use: it maintains that neither researchers nor agents saw Buckmaster and Alpöge's work by any means before it was published. But it adds a caveat: "although it is unlikely, we cannot rule out that de-identified data derived from your use of our products helped improve our models." As for why they took on Navier-Stokes so suddenly, Bubeck explained at a press conference that the team had heard rumors "on Twitter" about other researchers making progress on Millennium Problems and decided to try as well, without knowing at first that those rumors referred precisely to Buckmaster and Alpöge.

The mathematical community's reaction has been one of alarm, regardless of whether the most serious accusations are confirmed. Abhishek Saha, of Queen Mary University of London, describes OpenAI's conduct as "the kind of thing mathematicians don't normally do." Matthew Ballard, of the University of South Carolina, notes that "mathematics relies heavily on an informal norm of trust": sharing incomplete ideas with colleagues on the assumption that they will not turn into a competitive race. Jeremy Avigad, of Carnegie Mellon, admits that "the very idea that AI systems could steal ideas from our queries is chilling." Brendan Hassett, of Brown, points out that "given AI companies' track record of taking copyrighted work without permission or payment, it is natural for people to ask these questions." And Yang-Hui He, of the London Institute for Mathematical Sciences, fears mathematics could return to a state of secrecy resembling the era of patronage by families such as the Medici.

Our reading is that this story mixes two things worth keeping separate: a real capability and a public relations operation. That a model solved in 88 hours a problem that had been open for nearly 90 years is a genuine milestone in the application of AI to mathematics, and it fits what we have already been observing about AI-assisted science: there is a rung of superhuman "instrument" already demonstrated (AlphaFold and its Nobel being the precedent), a rung of "copilot" where results tend to be oversold, and a rung of "autonomous scientist" that no one has yet verified. This case, announced via a blog post and without the Clay Institute having awarded the prize, sits somewhere in between: it is a real technical achievement, but calling it "solved" before the community confirms it is, in itself, an exercise in framing.

The short-term cost here is neither speculative nor distant: it is the erosion of the informal trust that has allowed mathematics to advance by sharing half-baked ideas. If any clue left in a conversation —or in a session with a coding assistant such as Codex— can, by OpenAI's own admission, indirectly "help improve" its models, the rational response for any researcher is to stay quiet until later. That is exactly the opposite of what science needs in order to accelerate with AI: more openness, not less. The paradox is uncomfortable but real, and Andras Juhasz, of Oxford, captures it well when he calls this episode a public relations victory that could prove pyrrhic: if labs with almost unlimited resources can throw "10,000 mathematicians" at any problem the moment they detect someone else making headway, the logic of publishing openly breaks down for the very profession AI says it wants to help. In the long term we remain convinced that these tools can multiply the pace of scientific discovery and bring us closer to goals such as accelerating cures for disease; but that future requires labs to earn the research community's trust, not frighten it into silence.