OpenAI claims a proof of the Navier-Stokes existence and smoothness Millennium Prize problem

OpenAI says an internal AI system — described as significantly more capable than its flagship GPT-6 Astra — produced an analytical proof and a machine-verified Lean formalization showing that a smooth, finite-energy three-dimensional fluid flow can spontaneously develop a singularity in finite time under the Navier–Stokes equations, addressing one of the six unsolved Clay Mathematics Institute Millennium Prize questions.

What's new

OpenAI's post, "On the Navier–Stokes Millennium Prize Problem," says the result resolves statement "C" (and also "D") of the Clay Institute's official problem formulation. The company describes the collapse mechanism as a vortex that "spirals inward and gets increasingly elongated" until it becomes singular.

Specifics from the write-up:

• The attempt used a system of roughly 10,000 concurrent coordinating agents devoted specifically to the Navier–Stokes problem.
• The agents "arrived at their resolution on Saturday, September 5, about 88 hours after the first agents were launched."
• Mechanically verifying the proof as a Lean formalization took an additional 17 hours, run through GPT-6 Astra.
• Across every open problem the broader agent system attempted, it sent 4.9 million messages and used roughly 300 billion output tokens.
• OpenAI explicitly waives the associated prize money — "We do not intend to claim the Millennium Prize for this result" — and credits concurrent, independent work by mathematicians Levent Alpöge and Tristan Buckmaster, saying it recognizes "the priority of their work on forced Euler."

Context

The Clay Mathematics Institute listed seven Millennium Prize Problems in 2000, each carrying a $1 million award for a correct solution; only one, the Poincaré conjecture, has been resolved and formally accepted, by Grigori Perelman in the mid-2000s. OpenAI frames the underlying fluid-dynamics question — can smooth 3D motion ever break down — as unresolved "for roughly 90 years."

The claim is unrelated to OpenAI's August 1 announcement that an earlier internal Astra model had separately solved ten different open problems across sphere-packing, coding theory, and group theory; none of those involved fluid dynamics or the Clay Institute's prize list. It also follows Anthropic's early-September claim that Claude produced a computer-checked proof of Fermat's Last Theorem — a sign that frontier labs are increasingly racing to attach their models' names to famous, previously-unsolved problems in pure mathematics rather than benchmark leaderboards.

Why it matters

A verified resolution of a Millennium Prize Problem would rank among the most significant results in the history of mathematics. Publishing a Lean formalization is a real step beyond a natural-language claim, since Lean mechanically checks each inference in a proof — a higher bar than prose alone. But mechanical verification of a formalization doesn't by itself confirm that the underlying proof strategy is airtight or that its framing of the problem is uncontested; results of this scale typically face months or years of independent scrutiny from the broader mathematical community before wide acceptance, which is likely part of why OpenAI is explicit that it isn't claiming the associated prize money. Regardless of how that scrutiny plays out, it's among the most ambitious claims yet of an AI system contributing directly to an unsolved, named problem in core mathematics rather than a constructed benchmark.