OpenAI Says AI Solved the Navier–Stokes Millennium Problem: What the Proof Means for Mathematics

Posted on 11th Sep 2026 06:05:38 in Artificial Intelligence, Machine Learning

Tagged as: AI, OpenAI, Mathematics, Navier-Stokes, Machine Learning

On 8 September 2026, OpenAI announced that an internal AI system had produced a full solution to the Navier–Stokes existence and smoothness problem — one of the seven Millennium Prize Problems selected by the Clay Mathematics Institute in 2000, each carrying a US$1 million bounty. The result, accompanied by a preprint and a machine-checked formalization in the Lean proof assistant, would mark the first time a major open problem in pure mathematics has been cracked by a computer, if the mathematical community confirms it after review.

A 90-Year-Old Question About How Fluids Move

The Navier–Stokes equations apply Newton’s second law of motion — force equals mass times acceleration — to fluids. They treat a liquid or gas as a continuous medium rather than tracking individual molecules, and they are the workhorse behind aircraft design, weather forecasting, ocean-current modelling and the study of blood flow. The equations trace back to the nineteenth-century work of Claude-Louis Navier and George Gabriel Stokes.

For nearly a century, however, one question has resisted every attempt to answer it. In 1934, French mathematician Jean Leray proved that solutions to the equations exist in a generalized sense. But whether those solutions always stay smooth — or whether a perfectly ordinary, smoothly starting flow can blow up, developing infinite speed in a finite amount of time — remained open. In 2000, the Clay Mathematics Institute named this the Navier–Stokes existence and smoothness problem and put a US$1 million prize on a correct resolution.

The stakes are physical as well as mathematical. A real fluid cannot move infinitely fast. If the equations allow a singularity to form, then under certain conditions the continuum approximation itself breaks down, and the equations are no longer a faithful mirror of reality — beyond that point, modelling would need to descend to the level of individual particles.

The Claimed Result: A Vortex That Unravels

OpenAI says its system produced an analytical proof, together with a Lean formalization, showing that an initially smooth fluid at rest can develop a singularity in finite time — even while a smooth external force is applied and the fluid’s energy remains finite throughout the dynamics. In the official formulation of the problem by Charles Fefferman, this establishes the “breakdown” options, statements C and D: the answer to the Millennium question, according to OpenAI, is that smooth solutions can indeed break down.

The construction is a vortex — a spinning swirl of fluid that spirals inward and stretches outward, growing elongated “like spaghetti,” as the OpenAI team describes it. The central region shrinks while its speed grows without bound, yet the total energy stays finite. The technical heart of the proof is that the competing terms in the equations — acceleration, pressure gradients, momentum transfer and viscosity — must all become large while cancelling in a precise balance, leaving only a smooth external force even as the velocity explodes.

“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,” OpenAI computer scientist Ven Chandrasekaran told reporters at a press briefing. Because such behaviour is physically impossible for a real fluid, he added, it suggests that under certain circumstances the equations may not be a reliable model of physical reality.

Crucially, the proof is not merely asserted — it was formalized and machine-checked in the Lean proof assistant, and OpenAI has published both the paper and the Lean repository. That verification layer matters: human review of a proof of this complexity typically takes months or years, and a machine-checked formalization lets independent mathematicians inspect the argument at the level of formal logic. OpenAI has also said it does not intend to claim the US$1 million Millennium Prize for the result.

How It Was Found: 10,000 Agents and 300 Billion Tokens

The scale of the computation is unlike anything previously reported for a mathematics problem. According to OpenAI, the effort ran on an internal model that it describes as significantly more capable than GPT-6 Astra, with training that began on 28 August. The company launched the project on 1 September after hearing rumours that two Millennium Prize problems had been resolved.

The system was organized as a swarm of coordinating agents powered by that internal model, with tools including the ability to read a cached copy of the internet and to run code. The agents were subdivided into groups that could communicate internally, and different groups were prompted with different variants of each problem — for Navier–Stokes, some groups were steered toward proof-style variants and others toward disproof-style variants, to avoid anchoring the system to a single answer.

The breakthrough came in stages:

• Euler first: nearly 100 agents worked for about 50 hours to resolve a simpler related question — the regularity problem for the Euler equations, which are the zero-viscosity limit of Navier–Stokes — in its unforced variant. That success convinced the team that Navier–Stokes itself was within reach.
• The full problem: OpenAI scaled up to roughly 10,000 concurrent agents and shifted resources away from the other Millennium Problems. The agents arrived at the Navier–Stokes resolution on 5 September, about 88 hours after the first agents were launched.
• Formal verification: Lean formalization and verification took an additional 17 hours, carried out using GPT-6 Astra.

Across all the problems attempted, the agents exchanged 4.9 million messages and used roughly 300 billion output tokens — an amount TechCrunch estimated at about US$22.5 million of compute at Astra’s then-current rates. The Navier–Stokes portion alone involved 2.7 million messages and approximately 130 billion output tokens. Groups were periodically cross-pollinated, with Codex consolidating the most useful insights from each group so the follow-up work could build on the agents’ own intermediate results.

A Crowded Race — and a Bitter Dispute Over Priority

OpenAI was not alone at the finish line. On 7 September, one day before OpenAI’s announcement, Harvard mathematician Levent Alpöge — who is also an Anthropic employee — and NYU professor Tristan Buckmaster released a paper solving a version of the problem for the Euler equations with a smooth force, using a mix of Anthropic’s Claude and OpenAI’s Codex and Astra models. They also said a solution to the more general problem would be released soon. On the same day, Anima Anandkumar of Caltech and collaborators released a solution to the zero-viscosity Euler problem using a physics-informed neural network.

But the parallel effort quickly turned contentious. Buckmaster published a statement alleging that information about his and Alpöge’s progress had been passed to OpenAI, and that OpenAI’s answers about when its work began grew evasive under questioning. “It emerged that an entire team had been working on the problem,” Buckmaster wrote, “and that an insane amount of compute had been used.” He said it was eventually agreed that OpenAI’s first prompt had been sent only days earlier — after word of their work reached OpenAI. Buckmaster also alleged that OpenAI mathematician Sébastien Bubeck asked him to remove Alpöge’s credit as part of a proposed compromise, and that when Buckmaster pushed to make the dispute public, Bubeck replied, “Why would you ruin your career?” followed by “If you don’t want me to be nice, then I don’t have to be nice.”

OpenAI’s own post acknowledges the timeline: the effort began on 1 September after hearing a rumour that later turned out to relate to Alpöge and Buckmaster, and the company says it reached out on 6 September, after completing its Lean verification, to offer a concurrent release and a joint announcement recognizing their priority. OpenAI states that its researchers and agents “did not see any of their work through any means until they released it publicly,” and that no specific user data was accessed. It adds, however, that it “cannot rule out” that de-identified data derived from user interactions helped improve its models, while noting the proofs differ significantly — the Euler results are forced versus unforced. On 10 September, OpenAI updated the post with findings from an internal investigation into whether user inputs could have influenced the result.

What It Means for Mathematics — and for AI

The announcement drew measured but striking reactions from working mathematicians. “It is certainly an exciting day, as we contemplate the announcement of major advances in the human understanding of mathematics,” said Martin Bridson, president of the Clay Mathematics Institute. Luis Martínez Zoroa of CUNEF University in Madrid, whose own work underpinned parts of the approach, called it “a truly remarkable result.” OpenAI mathematician Sébastien Bubeck described it as “the spectacular culmination of the arc we have seen over the last 12 months.”

Yet the result is a claim until the wider community verifies it — and the mathematical world has learned to be careful with headline proofs. The Lean formalization is a genuine safeguard, because a machine-checked proof removes the single most fragile step in the review pipeline. Independent researchers will now scrutinize whether the construction truly satisfies every hypothesis of the official problem statement, particularly the smooth forcing and finite-energy conditions.

Beyond the prize, the episode sharpens two debates. The first is scientific: if verified, the theorem tells physicists that the continuum description of fluid motion has a hard boundary — at sufficiently extreme conditions, the equations that design aircraft and predict weather mathematically permit behaviour no real fluid can exhibit, and modelling must switch to finer-grained descriptions. The second is about credit: when a model trained partly on the accumulated work of a community produces a breakthrough, who deserves authorship, and how should priority be established when compute budgets — not just insight — can decide the race? OpenAI’s own framing is that the milestone is “not a culmination, but rather a snapshot in time” of AI progress, and the company says it plans to use what it learned to guide the pace of future capability advances.

For mathematicians and AI researchers alike, the Navier–Stokes announcement is a preview of a new mode of research: frontier models running tens of thousands of coordinated agents, spending hundreds of billions of tokens on a single question — and delivering a proof that now must survive the scrutiny of human minds.

Sources

• Nature — OpenAI claims huge maths breakthrough on a famed ‘Millennium Problem’
• OpenAI — On the Navier–Stokes Millennium Prize Problem
• TechCrunch — OpenAI fought dirty on career-making math problem, says NYU mathematician
• The Guardian — OpenAI claims to have solved maths problem that stumped humans for decades
• The New York Times — OpenAI Says It Has Cracked One of Math’s ‘Millennium Problems’