OpenAI Claims 10,000-Agent System Solved Navier-Stokes Problem in 88 Hours; Mathematician Disputes Independence of Resul

OpenAI says a system of approximately 10,000 coordinating AI agents produced a solution to the Navier-Stokes equations — one of the seven Millennium Prize Problems — in 88 hours, according to a release the company published Tuesday. The claim has not been verified by the Clay Mathematics Institute, which established the $1 million prizes in 2000 and has not yet commented on OpenAI's proposed proof.

According to OpenAI, the agents were powered by an unnamed internal AI model and organized into communicating subgroups, with the group that reached the Navier-Stokes result comprising roughly 10,000 concurrent agents. OpenAI says the agents had access to a cached version of the internet and could run code, and that the resolution was reached on Saturday, September 5 — about 88 hours after the first agents launched. The company has not disclosed the model's name, parameter count, architecture, or any technical specifications for the system used.

The Navier-Stokes equations describe the motion of viscous fluids and underpin fields from aerodynamics to weather modeling. The problem asks whether smooth, well-behaved solutions to the equations always exist in three dimensions, or whether singularities can form. It is one of seven Millennium Prize Problems; only one, the Poincaré Conjecture, has been solved since the prizes were announced in 2000.

Mathematician disputes independence of the work

Tristan Buckmaster, a mathematics professor at New York University, publicly challenged the claim on his website. Buckmaster said he had been collaborating personally with Levent Alpöge, a mathematician at Anthropic, on Navier-Stokes and related problems. He said Alpöge received tips that information about their progress had reached OpenAI, and that OpenAI's route to a solution resembled their own work in ways he says would be unlikely to emerge from a model given only the problem statement in a few days.

Buckmaster raised the possibility that OpenAI models had been trained on, or had access to, his and Alpöge's sessions using OpenAI's Codex tool, which the pair had used as part of their research process. He was careful to qualify his statement: "I have not seen OpenAI's proof. I do not know what their model did, or how. I do not know whether our data was used."

OpenAI responded in its release, saying its effort began September 1 after "hearing a rumor" about progress on the problem that it later traced to Alpöge and Buckmaster. The company said neither its researchers nor its agents saw any of the pair's work through any means until it was released publicly, and that "no specific user data was accessed in order to solve this problem." OpenAI added: "While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models."

What this means

No independent body has verified OpenAI's proof, and the Clay Mathematics Institute — the sole arbiter of Millennium Prize claims — has not weighed in. Until a peer-reviewed verification occurs, this remains an unverified claim rather than a confirmed mathematical result. The dispute with Buckmaster adds a second layer of uncertainty: even if the proof is correct, questions about whether it was derived independently or informed by data from a rival researcher's tool usage go to the heart of how AI labs handle user data from coding and research tools. OpenAI's acknowledgment that de-identified data "could" have influenced model training, even while denying direct access, is likely to fuel scrutiny of data provenance in AI-assisted research going forward. For now, treat the 88-hour timeline and the underlying proof as an OpenAI claim awaiting mathematical and institutional confirmation.