On 1 September this year, OpenAI engineers had heard a rumour: two of the Millennium Prize problems — the hardest unsolved questions in mathematics, each carrying a $1 million reward — may have been cracked by human researchers working elsewhere. Rather than wait, the company set roughly 10,000 AI agents to work on some of those same problems. By 5 September, just 88 hours later, it claimed a result that has eluded mathematicians for nine decades. The target was the Navier-Stokes existence and smoothness problem, a set of equations that describe how liquids and gases move.

Turbulence — the chaotic, unpredictable churning of fluids that makes aircraft shake and rivers surge — sits at the very heart of these equations, and it remains one of the least understood phenomena in physics. OpenAI says its solution resolved two of the problem's four required components. To reach that point, the AI agents exchanged nearly three million messages and generated 130 billion output tokens, a volume of computational effort that would have cost approximately $10 million if billed at the company's standard rates. The model that powered this effort is not yet available to the public.

Trained from late August, it is described by OpenAI as significantly more capable than any of its released products — a detail that will sharpen debates about how quickly the company should be sharing its most powerful tools with the world. What the result does not yet have, however, is independent confirmation. The Clay Mathematics Institute, which administers the Millennium Prizes, has not verified or publicly accepted OpenAI's claimed solution. In mathematics, a proof that has not been checked by the broader community is, in a precise and important sense, not yet a proof at all — and the distance between those two states is where the real story may still be unfolding.