OpenAI has announced a solution to the Navier-Stokes problem — one of mathematics' seven Millennium Prize Problems, unsolved for approximately 90 years, and worth $1 million to whoever gets there first. The AI got there first. The humans are still discussing what that means.
The announcement arrived Tuesday. The controversy arrived Monday.
OpenAI says it cannot rule out that de-identified data from the researchers' own sessions helped solve the problem they were trying to solve.
What happened
OpenAI began training an internal AI model on August 28th — more powerful than the recently released GPT-6 Astra, running 10,000 concurrent agents — and pointed it at the Navier-Stokes equations, which describe the behavior of fluid and gas flow. The model produced a solution. This took eleven days.
One day before OpenAI's announcement, NYU mathematics professor Tristan Buckmaster and Anthropic researcher Levent Alpöge published findings on a closely related problem. Buckmaster subsequently noted that the route OpenAI's model had taken looked, to an informed eye, rather familiar. It was the same route he and Alpöge had been developing — using OpenAI's own Codex, into which they had been entering their drafts throughout the project.
Buckmaster contacted OpenAI. He asked whether the model had been trained on their Codex sessions. He received a partial answer. He asked again, specifically about training data. He did not receive an answer to that one.
Why the humans care
The Navier-Stokes problem is not a curiosity. Its solution has implications for physics, engineering, climate modeling, and the fundamental mathematical description of how matter moves through space. A $1 million prize exists because, for nine decades, no human could collect it. The machine apparently could, and did, in under a fortnight.
The secondary concern — whether OpenAI's model benefited from the unpublished work of researchers who trusted OpenAI's own tools — is the kind of question that tends to matter more in the short term than the mathematical result itself. OpenAI's statement confirms it found no specific user data access, while noting it cannot rule out that de-identified data from those very sessions improved the model. This is a careful sentence. It was written by people who are good at careful sentences.
What happens next
OpenAI's proof will be reviewed by the mathematics community, which has its own timeline and does not particularly care about press releases. The $1 million prize awaits formal verification.
Somewhere, a mathematician is opening Codex to begin working on the next Millennium Problem. The irony is structural at this point. Welcome to the next step.