OpenAI Proves Navier-Stokes Can Blow Up, Then Walks Away

OpenAI said on September 8 that an unreleased model proved a smooth 3D fluid can blow up in finite time. The company posted a writeup and a Lean check, then said it will not claim the Clay Mathematics Institute’s $1 million Navier-Stokes prize.

The proof uses a smooth outside force. That is the breakdown option Clay wrote into the official problem, not the unforced smoothness question that still sits beside it.

OpenAI Will Not Collect the Million

Clay put seven problems on a list in 2000 and set aside $1 million for each, a $7 million fund with no deadline. Until this week the Poincaré conjecture was the only one with a widely accepted solution. OpenAI now says it has closed Navier-Stokes by showing a singularity, and in the same breath it is walking away from the money.

“Our goal in releasing this result is to report on the substantial progress of our AI models,” the company wrote. “We do not intend to claim the Millennium Prize for this result.” The post is a capability note as much as a math paper: the work ran on an internal model more capable than GPT-6 Astra, a system OpenAI has not shipped, whose training started on August 28 and is still running.

Venkat Chandrasekaran, a computer scientist at OpenAI, put the claim in plain language on a press call. “Our proof does show that there exist fluids which start out perfectly normal, and under the Navier-Stokes equations, actually achieve infinite speed in a finite amount of time,” he said.

THE MONEY AROUND THE PROOF

• Clay prize: $1 million for a solution that survives Clay’s own process.
• Company spend: Mark Chen, OpenAI’s head of research, said the run cost “in the millions of dollars.”
• Customer price: OpenAI told reporters a customer-priced rerun of the same job would cost about $15 million.
• Sebastien Bubeck: “We are able to spend millions of dollars on a problem that we really care for,” the OpenAI researcher said on the same briefing.

The prize is smaller than the sticker on the compute. That gap is the point of the release: OpenAI is showing what it can buy in 88 hours, not filing for Clay’s check.

The Force Clay Wrote Into Statement C

Charles Fefferman of Princeton wrote the official problem for Clay. He gave four acceptable targets. Two ask for global smooth solutions with no outside force. Two ask for a breakdown, and those two allow a smooth force.

FEFFERMAN’S FOUR TARGETS

OpenAI says it established statements C and D in Fefferman’s text. The fluid begins at rest. A smooth force is applied. Energy stays finite. Speed still grows without bound in finite time. That is a legal Clay target. It is also the version that lets someone push on the fluid from outside, which is why some analysts treat C and D as a different question from A and B.

The unforced regularity problem, the one with f set to zero, is the one aircraft codes and weather models most resemble. Nobody has posted a Lean-checked proof of A or B. OpenAI did not claim those.

Ten Thousand Agents Built a Spaghetti Vortex

The construction is a vortex that spirals inward and stretches, “like spaghetti,” OpenAI wrote. The core shrinks and speeds up while the energy stays finite. Acceleration, pressure, momentum transfer, and viscosity all get large and then cancel with enough care that the outside force remains smooth. The company says the breakdown has to come from the fluid’s own motion, not from an infinite shove put in by hand.

GPT-6 Astra did not find that object. OpenAI says an unfinished internal model did, after the lab heard rumors on Tuesday, September 1, that two Millennium problems had fallen. Agents got a cached web, a code runner, and variants A through D. Groups could talk inside the group. After an Euler surprise, the lab pulled people off the other prize problems and fed the Euler writeup forward.

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.

The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.

The problem… pic.twitter.com/8zol3BPTL4

— OpenAI (@OpenAI) September 8, 2026

Bubeck said he thought there must be a mistake somewhere, then had a Lean-checked argument on Sunday morning. The agents hit a resolution on Saturday, September 5, about 88 hours after the first launch. Astra spent another 17 hours on the Lean pass, finished on September 6. The Navier-Stokes folder and the Euler folder now sit together as Lean 4 certificates for both blowups, built on Lean 4.34.0-rc2.

THE NAVIER-STOKES RUN

• Agent count: On the order of 10,000 concurrent agents on the group that produced the Navier-Stokes result.
• Time to a proof: About 88 hours from the first launch on September 1 to the September 5 resolution.
• Messages: 2.7 million on Navier-Stokes, 4.9 million across every problem the agents tried.
• Output tokens: About 130 billion on Navier-Stokes, about 300 billion on the full attempt.

A 10,000-agent search at that token volume is test-time compute at industrial scale. It is not a single mathematician sitting with a pad. The lab used Codex to merge the better intermediate ideas from rival groups, then updated the agents when a newer checkpoint of the internal model came in.

Alpöge and Buckmaster Beat Them to Forced Euler

The rumor that started the sprint was not, in the end, a Navier-Stokes proof. OpenAI later tied it to Levent Alpöge, a mathematician employed by Anthropic, and Tristan Buckmaster, a math professor at New York University. After Lean verification on September 6, OpenAI reached out, still thinking the pair had Navier-Stokes, and offered a joint announcement. The pair had a forced Euler result, plus blowups for the incompressible porous medium equation and 2D Boussinesq.

OpenAI’s own Euler result is the harder-looking cousin: unforced, no outside push, viscosity stripped out. Nearly 100 agents spent about 50 hours on that disproof, the company said, before the lab decided Navier-Stokes was the better target. The two Euler theorems are not the same object. Forced Euler is Alpöge and Buckmaster’s priority, and OpenAI said so in print.

It is now widely expected that it should be possible to construct smooth initial data and smooth forcing term that would make these equations develop singularities in finite time; and it should even be possible to do without the forcing term.

Terence Tao, UCLA mathematician, 7 September 2026

Tao was writing about the Alpöge-Buckmaster papers, a day before OpenAI posted. He traced the iterative method to Diego Córdoba and Luis Martínez-Zoroa, whose Córdoba and Martínez-Zoroa blowup method adds high-frequency corrections until a solution breaks while the force stays tame. Fefferman, who wrote Clay’s text, said he was thrilled the problem was solved and named Córdoba and Martínez-Zoroa as the heroes of the story. Buckmaster has said Martínez-Zoroa deserves a Fields Medal.

THE WEEK THE PROOFS LANDED

1. August 28, 2026: OpenAI starts training the internal math model, still unreleased.
2. September 1, 2026: Rumors of prize-level fluid work reach the lab; agent groups launch on the open Millennium problems.
3. September 5, 2026: Agents produce a Navier-Stokes resolution, about 88 hours after launch.
4. September 6, 2026: GPT-6 Astra finishes Lean verification; OpenAI contacts Buckmaster.
5. September 7, 2026: Tao posts on the Alpöge-Buckmaster forced Euler, Boussinesq, and porous-medium blowups.
6. September 8, 2026: The pair’s papers go up; OpenAI publishes its Navier-Stokes claim the same day.

The human papers, Tao noted, still needed rewriting out of what the authors called the worst writeup they had seen. They posted early because of the OpenAI contact, not because the exposition was done.

The Hedge OpenAI Wrote About User Data

Buckmaster and Alpöge had been putting drafts through OpenAI’s Codex, and through Anthropic’s Claude, for a project they describe as independent of either employer. Buckmaster’s September 8 statement says he emailed OpenAI on September 3 to kill a rumor that Anthropic had solved a prize problem, then took two calls on September 6 with Bubeck. He writes that he was told an internal model already had a forced Navier-Stokes blowup on R³ and on the torus, Fefferman’s C and D, along the same smooth-force route Córdoba’s group had opened and that he and Alpöge had chosen.

I have not seen OpenAI’s proof. I do not know what their model did, or how. I do not know whether our data was used. I am not accusing anyone of anything. I am stating what I was told, when, and what was proposed to me.

Tristan Buckmaster, professor of mathematics, New York University, 8 September 2026 statement

He also wrote that Bubeck twice wanted Alpöge left off a joint authorship, and that when Buckmaster said he would go public, Bubeck asked, “Why would you ruin your career?” Bubeck called the circulating allegations “false and inflammatory.” He denied asking to drop Alpöge and said he had wanted to offer the pair the internal model so they could try to bridge Euler to Navier-Stokes, but did not know how to give an Anthropic employee access to OpenAI’s internal IP. OpenAI called the broader racing claims categorically false.

The company’s own post leaves a narrower door open. “We (the researchers and the agents) did not see any of their work through any means until they released it publicly, in particular, no specific user data was accessed in order to solve this problem,” it said. Then: “While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.” That sentence is doing a lot of work. Direct lookup is denied. Training-time leakage from Codex logs is not.

WHERE THE ACCOUNTS SPLIT

• On seeing the work: OpenAI says neither its people nor its agents saw the pair’s unpublished proofs; Buckmaster says the C/D route and the timing, after his rumor-killing email, look too close for comfort.
• On authorship: Buckmaster says Bubeck pressed to omit Alpöge because he works at Anthropic; Bubeck says that is false and that the hangup was access to an unreleased model.
• On the math: OpenAI says the Euler theorems differ (forced versus unforced) and the Navier-Stokes proofs differ; Buckmaster has not read OpenAI’s argument and is not claiming theft of a finished NS proof he did not have.

The first joke under OpenAI’s announcement was whether the solution was already in the training set. That is a crude version of the same question the company answered with a hedge rather than a no.

Why the Prize Clock Has Not Started?

Clay does not take manuscripts. It does not score blog posts, GitHub repos, or press briefings. A proposed solution has to appear in a qualifying outlet, meaning a refereed mathematics journal of worldwide repute with a named board, a real refereeing process, and a MathSciNet listing, unless Clay’s board later relaxes that bar. Then two years after a refereed journal paper it has to win general acceptance, as Clay alone judges it, before Clay even convenes a special committee.

WHAT CLAY REQUIRES BEFORE $1 MILLION MOVES

• Qualifying outlet: A refereed math journal of worldwide repute, not a company blog and not a direct submission to Clay.
• Waiting period: At least two years after that publication, with “rigorous examination” by the field.
• Acceptance: General acceptance in the global mathematics community, in Clay’s sole discretion.
• Either direction: For Navier-Stokes, a breakdown counts the same as a smoothness theorem, once those tests are met.

OpenAI is not starting that clock. It said so. Even a perfect Lean file does not replace a journal, and Lean does not decide whether C and D, with a force, are the problem Clay meant to buy. Clay also says it will pay special attention to whether a prize solution depends on earlier insights, and it may put prior authors on the citation or in the award. Córdoba, Martínez-Zoroa, Alpöge, and Buckmaster are already in that sentence, whatever happens to the million.

Infinite Speed Is a Model Failure

Navier-Stokes treats water or air as a continuum, not as a pile of molecules. Jean Leray showed in 1934 that weak solutions exist. Smoothness for 3D incompressible flow with constant density stayed open. A singularity in the sense OpenAI is using means speed becomes unbounded in finite time even though viscosity is trying to smear the motion out. A real fluid cannot do that. If the equations say it can, the continuum model has stepped outside physics, and the next description has to track particles.

That is why aircraft design, weather codes, and blood-flow models care. They already run on grids and cutoffs. A theorem that some smooth forced data explode does not make a 737 fall out of the sky. It says the ideal equation on R³ can leave the regime those codes pretend to inhabit. Unforced blowup, if anyone later proves it, would cut closer still. Tao, on September 7, thought that last step looked feasible. OpenAI says it has already done unforced Euler, and forced Navier-Stokes, in a week of agents.

The spaghetti vortex is a mathematical object with a machine-checked certificate. The $1 million is still in Clay’s fund. The model that wrote the proof is still inside OpenAI, and still training.