OpenAI's unreleased model solves a Millennium Prize maths problem: and the controversy may matter as much as the proof
10,000 AI agents, 88 hours, and a Lean-verified proof of finite-time singularity in Navier: Stokes. Here's what happened and why it matters.
On 8 September 2026, OpenAI published what may be the most significant mathematical proof ever produced with the help of artificial intelligence: a 165-page analytical proof, accompanied by a full Lean 4 formalisation, claiming to resolve the Navier, Stokes existence and smoothness problem. That is one of the seven Clay Millennium Prize Problems, each carrying a $1 million reward, and it has been open since 2000, though the underlying equations date back to the nineteenth century.
The result is remarkable. The controversy surrounding it is arguably just as important to understand.
What was actually proved
The Navier, Stokes equations describe how fluids move. Mathematicians have long known that approximate solutions exist, but a central question remained unanswered: do perfectly smooth solutions always stay smooth, or can they break down in finite time, developing what mathematicians call a singularity?
OpenAI’s proof establishes the latter. As OpenAI computer scientist Ven Chandrasekaran explained in a press briefing: “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.”
Physically, infinite velocity is impossible for a real fluid. The mathematical implication is significant: under certain conditions, the Navier, Stokes equations may simply stop being a reliable description of physical reality. The proof constructs a vortex, a spinning swirl of fluid that spirals inward and elongates, spaghetti-like, with its central region shrinking and accelerating while kinetic energy stays bounded. The technical achievement is making this breakdown emerge from the fluid’s own motion, not from any external forcing.
The proof resolves statements C and D of the official Millennium Prize formulation. OpenAI has published both the full 165-page proof and a Lean repository that independent researchers can download and verify themselves.
How the proof was made
OpenAI used an internal model described as “significantly more capable” than GPT-6 Astra, with training begun on 28 August 2026 and still ongoing. The model has no public name.
Starting on 1 September 2026, OpenAI deployed roughly 10,000 concurrent agents to work on the problem. These agents were organised into communicating groups with access to a cached internet snapshot and code execution tools. Different groups were assigned different provable or disprovable variants of the problem, and were actively encouraged to explore a diversity of approaches. As a further-trained version of the internal model became available mid-effort, agents were updated to it. At a later stage, Codex was used to cross-pollinate the most useful insights between groups.
Over 88 hours, the agents exchanged 2.7 million messages and consumed approximately 130 billion output tokens. GPT-6 Astra then spent an additional 17 hours formalising and verifying the proof in Lean 4. Computing costs ran into the millions of dollars.
As a side result, a group of nearly 100 agents spent around 50 hours resolving a related question: the regularity problem for the unforced Euler equations, the limit of Navier, Stokes with the viscosity term removed. The agents resolved this question without being specifically prompted to, which reportedly surprised the OpenAI team.
The controversy you need to know about
OpenAI launched its effort on 1 September 2026 after hearing a rumour that outside researchers were close to a Navier, Stokes breakthrough. Sam Altman has since confirmed the motivation publicly: “We were curious if ours could do it too.”
The researchers in question were Tristan Buckmaster, a mathematician at New York University, and Levent Alpöge, a researcher at Anthropic, working together on an independent personal collaboration. Over roughly a year, they had produced Lean-verified finite-time blowup results for several related fluid-dynamics problems, including 3D incompressible Euler with smooth forcing. By 22 August 2026, those results were in hand.
Buckmaster published his own announcement 12 hours before OpenAI’s, on 8 September 2026. He also made a more serious allegation: that OpenAI employees presented him with a choice. Either he and Alpöge could post their work and OpenAI would publish its Navier, Stokes solution the following day, or he could collaborate with OpenAI on the Navier, Stokes paper with Alpöge excluded from authorship, on the grounds of his affiliation with OpenAI’s biggest competitor.
Buckmaster also asked OpenAI whether its agents had accessed transcripts of work he and Alpöge had done using OpenAI’s own models. Employees denied that. When he asked whether those transcripts had been used in training, they did not respond.
OpenAI’s official statement acknowledges the question while stopping short of a full denial: “We cannot rule out that de-identified data derived from their usage of our products helped improve our models.” OpenAI mathematician Sébastien Bubeck stated in a press briefing that the internal model had independently solved the Euler problem by entirely different means.
OpenAI also clarified that when it reached out to Buckmaster and Alpöge on 6 September 2026 to offer a concurrent release, it was under the impression the pair had a Navier, Stokes solution. They did not. Their result was the forced Euler problem, a related but distinct achievement.
What does this mean for you?
If you work in or around mathematics, scientific research, or computational modelling, a few things are worth sitting with.
First, the pace. From rumour to Lean-verified proof took four days. That is a research tempo that simply did not exist two years ago. The Clay Mathematics Institute has not yet formally recognised the result, its rules require publication in a qualifying outlet, two years in print, and general acceptance by the community, but if the proof holds, the speed at which it arrived will be remembered as much as the proof itself.
Second, the physical implications are more modest than headlines suggest. The proof shows the equations can break down mathematically, not that fluids actually behave this way. For engineers and scientists using Navier, Stokes to model blood flow, weather systems, or aerodynamics, the equations remain as practically useful as they were before 8 September 2026. What changes is our theoretical understanding of their limits.
Third, and perhaps most pressingly: the competitive dynamic exposed here is uncomfortable. OpenAI provided the AI tools that Buckmaster and Alpöge used in their research. When those same tools helped surface the possibility of a breakthrough, OpenAI was in a position to mobilise vastly more compute to pursue the same result. The question of what happens to independent researchers when the infrastructure provider can also become a competitor is one the research community will be working through for some time.
OpenAI says it is not pursuing the $1 million prize. The organisation is framing this as a demonstration of how quickly its most advanced systems are improving. That is probably true. It is also, given the circumstances, a complicated kind of victory.