TL;DR
- On September 8, 2026, OpenAI announced that 10,000 coordinating AI agents, running on an unreleased internal model more capable than GPT-6 Astra, produced a 165-page Lean-verified proof of the Navier-Stokes existence and smoothness problem — one of seven Clay Millennium Prize problems open since 2000. - The agents sent 2.7 million messages and consumed roughly 130 billion output tokens over 88 hours. GPT-6 Astra then spent 17 additional hours formalizing the proof in Lean 4. Simon Willison estimated the compute cost at around $15 million at public API prices. - The result resolves statements C and D of the official problem formulation — a specific and significant subset. The Clay Mathematics Institute still lists the problem as unsolved, requires two years of post-publication community review, and OpenAI says it is not claiming the $1 million prize. - The story has a controversy attached. OpenAI launched the effort on September 1 after hearing a rumor that two Millennium Prize problems were close to resolution by outside researchers — specifically Levent Alpoge, a mathematician at Anthropic, and Tristan Buckmaster, a professor at NYU. OpenAI reached out to offer a joint release, but under the mistaken belief that the pair had a Navier-Stokes result. They did not. Their work was on a related but distinct problem. - The mathematical community is scrutinizing whether the Lean definitions faithfully capture the Clay problem statement, whether the proof depends on the right axioms, and what role training data may have played. None of those questions are settled.
The Navier-Stokes equations describe how fluids move. Write water in a bathtub. Write air around an aircraft wing. Write the atmosphere. If you can predict exactly how a fluid behaves under any conditions — smooth, turbulent, anything — you need to understand whether the equations that govern it can ever produce a singularity, a point where velocity becomes infinite. That question, whether smooth three-dimensional fluid flows must stay smooth forever or can blow up in finite time, has been open for about 90 years.
On September 8, OpenAI said it had an answer.
What OpenAI actually proved
The short version: OpenAI's agents showed that under certain conditions, the Navier-Stokes equations allow blow-up. A three-dimensional incompressible fluid, starting from rest, with a smooth external force applied to it, can develop unbounded velocity in finite time while its kinetic energy stays bounded and the forcing stays smooth.
That is statements C and D of the official Millennium Prize formulation. The Clay Mathematics Institute's problem description offers four statements — A through D — and says it is phrasing them this way to give solvers reasonable leeway while retaining the heart of the problem. Proving C or D counts.
OpenAI published the full 165-page analytical proof and the Lean 4 formalization on GitHub. Independent researchers can download it, run the build, and check it against the axioms. That is more than most mathematical proofs ever offer in the way of machine-checkable evidence.
The Clay Institute still lists the problem as unsolved. Its rules require publication in a qualifying outlet, two years of print time, and general acceptance by the mathematical community. Martin Bridson, Clay's president, said the evaluation would be deliberately unhurried and absolutely rigorous. OpenAI said it is not claiming the prize.
How 10,000 agents solved it in 88 hours
OpenAI began training the unnamed internal model on August 28. On September 1, its researchers heard a rumor that outside mathematicians were close to cracking two Millennium Prize problems. They deployed roughly 10,000 concurrent agents to work on the remaining open problems, giving different agent groups different formulations and different mathematical approaches, with access to a cached internet snapshot and code execution tools.
The agents communicated. 2.7 million messages over 88 hours, organized into groups exploring different constructive and non-constructive variants of the problem. When a further-trained version of the model became available mid-effort, agents were updated to it. OpenAI used Codex to consolidate useful intermediate results from different groups and redistribute them. By September 5, the Navier-Stokes groups had their result.
GPT-6 Astra then spent 17 hours writing the proof down in Lean 4 — not checking it, writing it. The Lean type-checker verifying a finished proof term takes minutes, not hours. Almost all of that 17 hours was a model translating a 165-page mathematical argument into a formal language a machine could verify. That is itself a significant capability. Human mathematicians typically describe Lean formalization as the most technically demanding part of formal proof work.
The compute cost, at public API prices, ran into what Simon Willison estimated as around $15 million. OpenAI's internal model is not publicly priced, so that is a proxy, not a bill.
The controversy
The rumor that triggered OpenAI's effort turned out to be about Levent Alpoge, a mathematician at Anthropic, and Tristan Buckmaster, a professor at NYU. The two had been working independently on related fluid dynamics problems. By August 22, they had Lean-verified finite-time blowup results for 3D incompressible Euler with smooth forcing.
When OpenAI completed its Lean verification on September 6, it reached out to them to offer a concurrent release, believing from the rumor that they also had a Navier-Stokes result. They did not. Their result was forced Euler — related, significant, but a distinct problem. The offers for joint announcement involved what one account described as commercial terms. The pair declined.
Buckmaster described one of the LLM-generated proofs from his own collaboration as the most horrendous he had ever read. The Lean verification means the formal statement is correct regardless. But it raises a genuine question: what does it mean for mathematics if we can produce verified proofs that are machine-correct but nearly unreadable by humans?
The harder question is about training data. OpenAI's statement acknowledges that it cannot rule out that de-identified data derived from Alpoge and Buckmaster's usage of OpenAI products helped improve the model. OpenAI says the researchers and the agents did not see any of their work through any means until it was released publicly. Alpoge and Buckmaster's result was on a different problem with a different approach, which makes direct copying implausible. But the ethics of whether a user's private mathematical work could improve a model that then wins a race against that same user are not resolved by the technical fact that the proofs are different.
What the Lean formalization actually tells you
The 17 hours of Lean formalization matters more than its runtime suggests. A Lean proof that compiles is mathematically certain in the following narrow sense: it follows from the axioms encoded in the type system, under the definitions used in the Lean file. Two things still require human judgment.
First, do the Lean definitions capture the Clay problem? Does the divergence operator in the Lean code mean classical divergence? Does the smoothness condition at t=0 include the one-sided derivative? Do the force-decay conditions match what Fefferman's official problem statement requires? These are questions about whether the formal theorem says what everyone thinks it says. A machine cannot answer them.
Second, what axioms does the proof depend on? Running the dependency checker should show at most three standard axioms. If it shows anything else — a sorry, a native_decide, a custom axiom — the proof may be bypassing the intended checks.
The Lean repository is public. Mathematicians are checking it. Whether the formalization faithfully captures the Clay problem is a question that the mathematical community will answer over months, not days.
Why the pace matters as much as the result
The most striking number is not 10,000 agents or 130 billion tokens. It is four days. From rumor to Lean-verified proof of a Millennium Prize problem: four days.
That tempo did not exist two years ago. It may not mean the problem is solved to Clay's standard — two years of community review is a real constraint, and there are genuine mathematical questions about the formalization. But if the proof holds, the speed will be remembered alongside the result.
Terence Tao, one of the mathematicians most involved in the existing foundational work the proof builds on, called the prior researchers the heroes of the story. The AI community is framing this as evidence of what coordinating agent systems can do at scale. The mathematical community is examining whether the definitions are right. The privacy community is asking about training data.
All of those conversations are happening simultaneously, which is roughly what you should expect when a swarm of 10,000 AI agents completes in 88 hours what humans have been attempting for nine decades.
The Clay Mathematics Institute will take its time. That is probably appropriate.
