OpenAI said an internal system solved the Navier-Stokes existence and smoothness problem on September 8, 2026, one of the seven Millennium Prize Problems named by the Clay Mathematics Institute. The writeup and a Lean formalization argue that a three-dimensional incompressible fluid, starting at rest with a smooth force and finite energy, can develop a singularity in finite time. That is a blowup: speeds grow without bound even though viscosity should smooth the flow. OpenAI said the result matches statements C and D in Clay’s official formulation, and that it does not intend to claim the $1 million prize.

The model that produced the proof is an unreleased system the lab said is “significantly more capable” than GPT-6 Astra. Training started August 28. After rumors on September 1 that two Millennium problems had been resolved, OpenAI pointed groups of coordinating agents at every open prize problem. The Navier-Stokes group ran on the order of 10,000 concurrent agents with a cached web reader and a code runner. Agents first surprised the team by disproving regularity for the unforced Euler equations (Navier-Stokes with viscosity removed). The lab then concentrated compute on Navier-Stokes. The agents reached a resolution on September 5, about 88 hours after launch. GPT-6 Astra spent another 17 hours on the Lean check. OpenAI said the Navier-Stokes run used about 2.7 million messages and 130 billion output tokens.

The constructed solution is a vortex that spirals inward and stretches like spaghetti, with motion, pressure, and viscosity terms growing large yet canceling so the external force stays smooth. OpenAI later realized the September 1 rumor concerned Levent Alpoge, an Anthropic employee, and Tristan Buckmaster of NYU, who had a result on forced Euler. The lab said it offered a joint announcement, recognized their priority on that Euler variant, and did not inspect their private work. It added that it cannot rule out de-identified product usage helping the models, while arguing the proofs differ.

Clay’s own problem page still listed Navier-Stokes as unsolved when this story was reported. A company blog post and a Lean file are not the Institute’s judgment.

Decoded Take

This is the loudest capability claim of the week, and it is also a process story. OpenAI did not wait for Clay, or for a journal, before telling the world that an unnamed model beat a 90-year question. The useful receipt is the Lean artifact, not the press line. If independent groups can replay the check and the argument survives, labs just moved from “AI helps write proofs” to “AI can finish a Clay-class disproof on a long weekend.” If the formalization has a hole, the post becomes a warning about racing rumors with 10,000 agents. Watch whether Clay updates the unsolved label, whether Alpoge and Buckmaster’s Euler paper and this Lean file get treated as a pair, and whether the next internal model is described by a paper rather than another countdown clock.