OpenAI solves a Millennium Prize problem in a week — and a mathematician working toward the same result accuses the company of playing dirty

OpenAI solves a Millennium Prize problem in a week — and a mathematician working toward the same result accuses the company of playing dirty

Aleksandr Manin
September 9, 2026, 09:05 AM

On September 8, OpenAI published a solution to the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems. It was produced by the company's internal model, which OpenAI describes as substantially more powerful than the recently released GPT-6 Astra.

OpenAI logo and the text GPT 6 Astra
OpenAI logo and the text GPT 6 Astra

The answer came out negative. The model proved that the three-dimensional Navier–Stokes equations can produce a singularity in finite time: the fluid's velocity at certain points goes to infinity, which doesn't happen in reality. In plain terms, the equations used to describe blood flow and weather fronts turned out to be fundamentally flawed.

The company doesn't intend to claim the million-dollar prize, for two reasons. The Clay Institute's rules require publication in a peer-reviewed journal and a two-year waiting period. On top of that, the economics don't add up: according to trade press estimates, a week of work burned through 300 billion tokens — roughly $22.5 million at Astra's current rates. Around ten thousand agents, according to company representatives, worked on the problem.

The scandal started a day before the announcement. NYU mathematician Tristan Buckmaster said that, together with Anthropic mathematician Levent Alpöge, he'd solved the forced Euler equations problem — a step toward Navier–Stokes, but not a full solution. They worked from an approach by Diego Córdoba and Luis Martínez-Zoroa, using Claude and Codex on the GPT-5.6 Sol model.

Visualization of a spiral fluid rotation
Visualization of a spiral fluid rotation

From there it's Buckmaster's version of events. According to him, OpenAI mathematician Sébastien Bubeck offered two options: either a joint publication in which the company would acknowledge the pair as the "people closest to the problem," or Buckmaster could publish alone and claim the prize — but only if he confirmed OpenAI's model's solution and removed Alpöge's name from the paper, since the company wasn't comfortable with his work at Anthropic. Buckmaster refused and warned he'd make the correspondence public. The reply, he claims, was: "Why would you want to ruin your career?" followed by, "If you don't want me to be polite, I don't have to be polite."

The mathematician is separately concerned about training data.

I asked whether the model had trained on our Codex sessions or had access to them, and I was told the model hadn't looked at user data. I asked again about training and got no answer.
— Tristan Buckmaster

OpenAI denies the accusations. According to the company, work began September 1 after hearing a rumor, verification in the Lean system finished September 6, and they only learned of their colleagues' results from the public announcement; no specific user data, they say, was used. That said, the company acknowledges two things: its model solved the full problem using a method similar to Buckmaster and Alpöge's approach, and indirect influence from anonymized data can't be fully ruled out. The Euler problem, OpenAI insists, the model reached by a completely different route.

An important caveat: the proof hasn't yet passed review by the mathematical community. Until independent confirmation, it's premature to call the Millennium Prize problem closed.

Do you think AI should even be eligible for math prizes — or should those stay in human hands?

    About the author
    Comments0