OpenAI said on September 8 that it has produced a solution to the Navier-Stokes existence and smoothness problem, one of the seven Millennium Prize Problems that have carried a $1 million award since the Clay Mathematics Institute set them in 2000. The problem, open for roughly 90 years, asks whether smooth three-dimensional fluid flow governed by the Navier-Stokes equations can always be extended indefinitely without breaking down.

What the agents found

According to OpenAI's writeup, the result was produced not by a single model call but by a coordinated group of roughly 10,000 AI agents, running on a next-generation model more capable than the company's publicly available GPT-6 Astra, that exchanged about 2.7 million messages and generated on the order of 130 billion tokens while working the problem. The agents identified a "finite-time blowup": a configuration in which a vortex tightens and spins increasingly fast in finite time while the fluid's total energy stays bounded, meaning smooth solutions do not always exist for all time in three dimensions. OpenAI said the model then spent roughly 17 more hours formalizing the argument, and the resulting proof has been checked in the Lean formal-verification language, giving outside mathematicians a mechanical way to confirm its logical validity independent of OpenAI's own claims. The company said it does not intend to claim the Clay Institute's prize money.

Reception is not unanimous

Coverage from Quanta Magazine and Nature notes that outside mathematicians are still working through the Lean-verified proof and treating the announcement as a claim to be independently confirmed rather than a settled result, given the problem's difficulty and the unusual scale of the automated effort behind it. Axios reported a separate controversy over credit, with questions raised about how much of the underlying mathematical insight originated with the AI system versus human researchers who guided or previously worked on related approaches.

Why it matters

Whatever its final reception among mathematicians, the announcement is a marker for how AI labs are now pitching large coordinated agent systems, rather than single chat-style queries, as a way to attack open problems that have resisted human effort for decades. It also lands during an unusually dense stretch of frontier-model competition, days after OpenAI's own GPT-6 Astra rollout and Anthropic's continued push into agentic tooling, as labs increasingly compete on what their systems can autonomously produce rather than only on benchmark scores.