OpenAI's Navier-Stokes claim clouded by rival math dispute

OpenAI says an internal model, more capable than its GPT-6 Astra system, has solved the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems the Clay Mathematics Institute has offered $1 million to resolve since 2000. The company announced the result on September 8, 2026. It is now facing pointed questions about how it got there.

The timing is the issue. One day before OpenAI's announcement, mathematician Tristan Buckmaster of New York University and Levent Alpöge, a researcher at rival lab Anthropic, published their own related breakthroughs, establishing finite-time blowup for the incompressible porous medium equation, the two-dimensional Boussinesq system, and the three-dimensional incompressible Euler equations.

Buckmaster and Alpöge had been developing their work inside OpenAI's Codex model, which raises an obvious question: did anyone at OpenAI see it before it went public? Sebastien Bubeck, an OpenAI researcher, denies the company used their results or accessed the material ahead of release.

OpenAI's own account complicates that denial. The company acknowledged it cannot rule out that de-identified data generated from Buckmaster and Alpöge's use of OpenAI's products "helped improve our models."

OpenAI acknowledged it cannot rule out that de-identified data "helped improve our models."

The Guardian reports that Buckmaster says he learned on September 3 that details of his work had reached OpenAI, and that he spoke directly with company representatives on September 6, two days before the public announcement. That sequence sits uneasily alongside Bubeck's assurance that OpenAI had no prior knowledge.

A separate discrepancy concerns when OpenAI's effort even began. An Arabic-language report citing OpenAI puts the training start date at August 28, 2026, while other accounts attribute to OpenAI a September 1 start date, a gap of roughly three days that has not been reconciled.

On the technical side, OpenAI's proof targets finite-time blowup for the forced Navier-Stokes equations, a distinct formulation from the equations Buckmaster and Alpöge addressed, meaning the two results are not directly overlapping even if they emerged from the same computational environment.

The scale of the compute involved is notable regardless of the dispute. GPT-6 Astra reportedly took about 17 hours to verify the proof once produced, while a swarm of roughly 10,000 AI agents working in parallel reached the solution itself in 88 hours.

Peer review of the proof has not been completed, and the mathematics community has not yet ratified the claim against Clay Institute standards. Until that happens, and until OpenAI clarifies exactly what its systems saw and when, the achievement remains as contested as it is impressive.