88 Hours, 130 Billion Tokens: Inside OpenAI's Navier-Stokes Proof

On September 8, 2026, OpenAI announced that an internal AI system had produced a solution to the Navier–Stokes existence and smoothness problem, one of the Millennium Prize Problems for which the Clay Mathematics Institute offers a $1 million award. The claim came with unusual operational detail. A new internal model, which OpenAI says is significantly more capable than GPT-6 Astra, coordinated a system of up to roughly 10,000 concurrent agents. That swarm exchanged 2.7 million messages and generated approximately 130 billion output tokens over about 88 hours to reach the resolution; GPT-6 Astra then formalized the proof in the Lean language in an additional 17 hours. OpenAI researcher Sébastien Bubeck estimates the computational cost at several million dollars, as reported by Quanta Magazine. Within a day, the announcement had also collided with a very human fight over who deserved credit for getting there — one with uncomfortable implications for any company whose confidential material flows through a vendor's AI tools.

What was actually claimed

The Navier–Stokes equations apply Newton's second law to fluids — water, air, blood — treating them as continuous media rather than tracking molecules. The Millennium Prize question is whether a three-dimensional incompressible fluid that starts out smooth can develop a singularity: a point where speeds grow without bound in finite time, despite viscosity's smoothing effect. The question has been open for roughly 90 years.

OpenAI's system produced an analytical proof and a Lean formalization showing that an initially smooth fluid at rest, driven by a smooth external force, can develop a singularity in finite time while its total energy stays bounded throughout. The solution is a vortex that spirals inward and elongates — OpenAI likens it to spaghetti — with the central region shrinking as it speeds up. The company says this establishes statements "C" and "D" of the Clay Institute's official problem formulation, the variants that permit smooth forcing.

Two caveats matter more than the headline. First, Clay's official page still lists the problem as unsolved as of September 10, 2026, and OpenAI states plainly that it does not intend to claim the prize — under Clay's rules, a solution must be published in a qualifying journal and survive roughly two years of scrutiny before any award is considered. Second, Lean verification confirms the logical chain, but as Quanta noted, humans must still confirm that the statement proven in Lean matches the mathematics being claimed.

The result is nonetheless being taken seriously. Charles Fefferman, the Princeton mathematician who wrote Clay's official description of the problem, told Quanta he was thrilled the problem was solved — while crediting Diego Córdoba and Luis Martínez-Zoroa, whose multi-year program of forced-blowup construction is the foundation both competing teams built on. If the result holds, Quanta assessed, it is by a significant margin the most important mathematical proof ever arrived at by an AI model.

How 10,000 agents found a proof

The methodology is as notable as the mathematics. OpenAI says it has been training a new internal model since August 28 that has shown unprecedented performance on its mathematics benchmarks. On September 1, after hearing rumors that two Millennium Prize problems had been resolved, the company launched an evaluation of the model against all open Millennium Prize problems.

The agents ran in groups that could communicate internally and had tools including reading a cached version of the internet and executing code. Different groups received different variants of each problem statement — for Navier–Stokes, versions A and B (which would require proofs of global regularity) and C and D (which would require blowup constructions) went to separate groups. First, on a set of "easier" problems, roughly 100 agents worked about 50 hours to disprove regularity for the unforced Euler equations. Seeing that, OpenAI shifted computing resources toward Navier–Stokes, used Codex to consolidate the most useful insights across agent groups, and swapped in a further-trained model mid-effort. The resolution arrived on Saturday, September 5 — about 88 hours after the first agents launched.

The scale deserves a sober reading. Spread across 10,000 agents, 130 billion tokens over 88 hours works out to roughly 40 tokens per agent per second — ordinary generation speed for a frontier model. What was extraordinary was the industrial coordination: sustained parallel exploration, cross-pollination of intermediate results, and formal filtering of errors. Across all attempted problems, the agents sent 4.9 million messages and used about 300 billion output tokens. This was not a flash of machine insight; it was a research factory.

The concurrent-work dispute

The same rumors that triggered OpenAI's effort had a source. Tristan Buckmaster, a mathematician at NYU, and Levent Alpöge, a researcher at Anthropic, had been running a yearlong personal collaboration — with no institutional involvement from either employer — using Claude, OpenAI's Codex, and other models. On August 15 they obtained blowup results with smooth forcing for the Boussinesq and incompressible Euler equations, verified in Lean on August 22. They delayed publication to rewrite machine-generated arguments into readable mathematics.

Buckmaster's account, published as a four-page statement just before OpenAI's announcement, describes what happened next. On September 3 he wrote to a prominent OpenAI mathematician to clarify the rumors. On September 6, after two phone calls with Bubeck, he says he was initially told an internal model had produced the proof with "very little human input" — and that over the course of the calls it emerged that a team had been working on the problem for days, had started with the unforced case, had first set the model on easier problems including Euler, and had used an enormous amount of compute. He says OpenAI eventually agreed the first prompt had been sent in the days after information about their work reached the company. Buckmaster also describes two publication proposals — including one in which he alone would write up OpenAI's Navier–Stokes result — and says Bubeck twice pushed to exclude Alpöge from authorship because he works at Anthropic. When Buckmaster said he would go public, the reply, per his statement, was: "Why would you ruin your career?"

OpenAI's blog tells a parallel story with different emphases. The company confirms its effort began September 1 after hearing a rumor it later realized related to Alpöge and Buckmaster; that it reached out after completing its own proof and Lean verification on September 6; that it offered visibility into all its prompts; and that it recognizes the pair's priority on forced Euler. On the central question, OpenAI states that neither its researchers nor its agents saw any of the two mathematicians' work before public release, and that no specific user data was accessed to solve the problem. But it adds a sentence every enterprise buyer should read twice: while unlikely, "we cannot rule out that de-identified data derived from their usage of our products helped improve our models." Buckmaster, for his part, is explicit that he is not alleging theft — "I am not accusing anyone of anything. I am stating what I was told, when, and what was proposed to me."

Why this is an enterprise story

Strip away the fluid dynamics and this is a story about what happens to sensitive material inside vendor AI systems. Two world-class researchers had spent a year pasting unpublished, career-defining work into commercial coding tools. The vendor's own carefully lawyered statement concedes it cannot rule out that such material, once de-identified, improved its models. Three practical takeaways follow.

Inventory what flows through vendor tools. Your researchers and engineers are pasting unreleased methods, unreleased code, and strategy documents into consumer-tier AI products today. That is the default surface where "data used to improve models" applies.

Contract for the gray zone. OpenAI itself offers Zero Data Retention for eligible API customers, under which prompts and responses are not retained after processing and enterprise customer data is not used for training unless customers explicitly opt in. Whether your agreements say something equivalent — and whether your most sensitive workloads run under it — is now a boardroom-level question.

Demand machine-checkable artifacts. What makes OpenAI's claim reviewable at all is that the Lean formalization is public, so any third party can verify the logical chain rather than trust the announcement. The same principle — formal checks, eval logs, reproducible evidence rather than press releases — is the right standard for consequential AI outputs inside your own shop.

Buckmaster closed his statement by calling this "a Deep Blue–Kasparov moment" that demands unhurried community discussion. For enterprises, the discussion is narrower but urgent: your most valuable material now moves through infrastructure owned by companies that are also your most capable competitors — for talent, for ideas, and increasingly for the work itself.

As of September 10, 2026, the Clay Mathematics Institute still lists Navier–Stokes as unsolved. The proof now enters the slow machinery of verification — and the faster machinery of commerce.