On September 8, 2026, OpenAI announced that an internal system of nearly 10,000 coordinating AI agents had produced a machine-verified proof that the Navier-Stokes equations for three-dimensional fluid motion can develop a singularity in finite time. This is one of the seven Clay Millennium Prize Problems, each carrying a $1 million bounty, and only the second to be solved in 26 years. The result matters not just for fluid dynamics, but for what it reveals about how frontier AI models can now tackle problems at the absolute frontier of human mathematics.
The Problem: Why Fluids Break the Math
The Navier-Stokes equations describe how fluids move using Newton's second law of motion. They govern everything from aircraft design and weather forecasting to blood flow in the human body. The equations treat a fluid as a continuous medium rather than tracking individual molecules, and the question at the heart of the Millennium Prize problem is whether this continuum approximation can break down.
Specifically, can the equations for a three-dimensional incompressible fluid with constant density develop a singularity - a point where calculated speeds grow without bound - even when the motion starts smoothly? If so, the mathematical model would fail to describe the real fluid, because real fluids cannot move infinitely fast. One would need to track the behavior of each particle individually.
The equations date to the nineteenth-century work of Claude-Louis Navier and George Gabriel Stokes. In 1934, Jean Leray proved that solutions exist in a generalized sense, but whether they always remain smooth became a central unanswered question. In 2000, the Clay Mathematics Institute named it one of seven Millennium Prize Problems.
As Charles Fefferman of Princeton University, who wrote the Clay Institute's official description of the problem, told Quanta Magazine: the fact that singularity formation is possible even in an idealized sense tells us that "straightforward as Newton's second law appears to be, its consequences when applied to fluids are profoundly counterintuitive. Turbulence is even weirder than it appears to be."
What the Proof Actually Shows
OpenAI's system produced an analytical proof and a formal Lean 4 verification that an initially smooth fluid at rest can develop a singularity in finite time under the influence of a smooth external force. The fluid's energy remains finite throughout the entire dynamics, from rest to the formation of the singularity.
The solution is a vortex - a spinning swirl of fluid that spirals inward and gets increasingly elongated, as OpenAI described it, "like spaghetti." This central region shrinks while it speeds up in such a way that its energy stays finite, as required by the laws of physics. The technical challenge is for the breakdown to occur through the motion of the fluid itself, not through an infinite force applied by hand.
The proof resolves the Millennium Prize problem by establishing statements "C" and "D" in the official Clay formulation. These are the "breakdown" variants, which ask for an example of a fluid that forms a singularity after being pushed by an external smooth force. This is distinct from the harder "existence and smoothness" question about unforced fluids - but it meets the Millennium Prize criteria.
As mathematician Nick McGreivy explained: the proof finds a velocity field u where u has a singularity, but the Navier-Stokes operator D[u] is smooth. The trick is to make the derivative terms in D[u] exactly cancel near the singularity such that the remainder is smooth. Setting f = D[u] gives a smooth forcing function. You never need to compute f explicitly.
The 10,000-Agent Swarm: How the Proof Was Found
The effort began on September 1, when OpenAI researchers heard rumors that Millennium Prize problems had been resolved. Using an internal model described as "significantly more capable than GPT-6 Astra," they launched a multi-agent system to evaluate all open Millennium Prize problems and several other high-impact problems.
The agents had access to tools including a cached version of the internet and the ability to run code. They were subdivided into groups that could communicate internally, with group sizes varying. The group that produced the Navier-Stokes resolution involved on the order of 10,000 concurrent agents.
Across all attempted problems, the agents sent 4.9 million messages and consumed approximately 300 billion output tokens. For the Navier-Stokes problem specifically, the agents sent 2.7 million messages and used roughly 130 billion output tokens. At current GPT-6 Astra rates of $10/$50 per million input/output tokens, that translates to an estimated $22.5 million in compute costs for the week-long effort.
The agents explored diverse approaches, and after some time, OpenAI used Codex to consolidate insights from each agent group - cross-pollinating results that drew on intermediate outputs. A key early result was the resolution of the unforced Euler equations (the Navier-Stokes equations with the viscosity term removed) by nearly 100 agents working together for approximately 50 hours. Once that result was achieved, OpenAI shifted agents away from other Millennium Problems and devoted resources to Navier-Stokes, guiding the agents with the Euler resolution as context.
The solution was reached on September 5, about 88 hours after the first agents were launched. Lean formalization and verification took an additional 17 hours using GPT-6 Astra. The 341,000 lines of Lean code ship in the open-source repository at github.com/openai/NavierStokesAndEuler with zero unproven axioms beyond standard Lean infrastructure.
The Lean Verification: Why Formal Proofs Matter
The repository ships machine-checked proofs in Lean 4.34.0-rc2 with Mathlib, using Lake for builds. The formalization covers two theorems for Navier-Stokes (one for Euclidean space R3, one for the periodic torus T3) and two for Euler. Each theorem file shows sorry_count: 0, meaning every step is formally verified.
Lean does not prove that the statement being formalized is the correct problem - only that the formalized statement is sound. As one commentator noted: "Lean guarantees the proof of a statement is sound, not that the statement is the problem you think it is. The first thing a checker confirms is whether what got formalized is actually Navier-Stokes with its viscous term, or something quietly weaker."
This is why human verification of the statement itself remains essential. The Millennium Prize criteria require a proof that meets specific conditions laid out in the Clay Institute's official statement, and that verification happens outside the formalizer. However, the Lean proof eliminates the most common failure mode of AI-generated math: errors in the proof chain itself.
For a result this significant, formal verification is not optional. It is the floor, not the ceiling. The mathematics community will need months to confirm that the formalized statement faithfully captures the Millennium Prize formulation, that the proof's construction is correct, and that the vortex singularity construction meets all technical requirements.
The Credit Dispute: Buckmaster, Alpoge, and OpenAI
The announcement came 12 hours after NYU mathematician Tristan Buckmaster posted three proofs on September 7 at 11:58 PM Eastern, just before the deadline. Working with Anthropic researcher Levent Alpoge, Buckmaster had used both Codex and Claude models to develop a similar approach to the problem. Their work was based on the techniques of Diego Cordoba of the Institute for Mathematical Sciences in Madrid and Luis Martinez-Zoroa of CUNEF University.
Buckmaster's statement revealed a tense sequence of events. After word of their progress leaked to OpenAI, they received calls from OpenAI's Sebastien Bubeck. According to Buckmaster, he was presented with two options: coordinate a joint release, or have Buckmaster alone present OpenAI's proof while leaving Alpoge off due to his Anthropic affiliation. When Buckmaster declined and said he would go public, Bubeck reportedly replied: "Why would you ruin your career?"
Bubeck later apologized for his "extremely poor choice of words," saying he was trying to ensure Buckmaster received proper credit and did not mean the statement literally. He also denied asking for Alpoge to be removed from authorship of his own work, saying he had suggested it would be inappropriate for an Anthropic employee to author OpenAI's work.
The core dispute centers on whether OpenAI's effort was inspired by or derived from Buckmaster and Alpoge's work. OpenAI confirmed it began on September 1 after hearing rumors, and that the team used a "smooth force" approach - the same niche route that Cordoba and Martinez-Zoroa pioneered and that Buckmaster and Alpoge were pursuing. As Buckmaster wrote: "Almost nobody else I know of was working on it. It is not the direction one arrives at in a few days by giving a model the problem statement."
OpenAI's statement acknowledged the possibility that de-identified data from Buckmaster and Alpoge's use of OpenAI products could have improved models, while maintaining that the proofs differ significantly and that the Euler results are different (forced vs unforced).
What Remains Open: The Harder Question
The Millennium Prize has been met, but the deeper mathematical question remains. OpenAI proved that forced Navier-Stokes can blow up. The existence and smoothness of unforced Navier-Stokes solutions - where no external force is applied - is still open.
The unforced case is strictly harder. As McGreivy explained: finding a solution where D[u] = f with f smooth was the trick in the forced case. For unforced equations, you need D[u] = 0 exactly, meaning the derivative terms must cancel perfectly with no remainder. There is no room for approximation.
There is reason to think this might be achievable. OpenAI simultaneously proved that unforced Euler equations (the zero-viscosity limit) can develop singularities. The remaining question is whether viscosity, which tends to smooth out motion, prevents the same breakdown in Navier-Stokes. If unforced Navier-Stokes blowup can be found, the entire problem will be settled.
As the Barcelona-based mathematician Cordoba noted: "Ten years ago, nobody believed there was a singularity for Navier-Stokes." The intellectual debt to the analytic techniques developed by Cordoba and Martinez-Zoroa is clear, as both competing AI-enabled groups built on their work. As Fefferman said: "The heroes of the story are Cordoba and Martinez-Zoroa."
The Bigger Picture: AI as Mathematical Engine
This result marks a turning point in how mathematical research can be conducted. A swarm of 10,000 agents consumed $22.5 million in compute to produce a proof that had resisted human mathematicians for decades. The proof ships as open-source code that anyone can verify. The formal verification is machine-checked and reproducible.
The implications extend beyond this single problem. Math proofs are largely verifiable through Lean, making them ideal reward signals for reinforcement learning. Anything that can be computationally verified can be used to train models toward superhuman performance in that domain. This creates a new paradigm for AI in science: systems that automate the discovery process itself, not just the analysis of data.
The credit dispute also reveals an emerging tension in AI-enabled research. When researchers use AI tools to develop unpublished ideas, they need clear guarantees that their work will not inform competing efforts by the companies whose tools they used. As Melissa Lee of Monash University said: "Researchers need to know very clearly what happens to unpublished work they enter into AI systems."
In the 26 years of the Millennium Prize, only one problem had been solved before - the Poincare conjecture in 2003, by Grigori Perelman. Now two have been resolved, both with AI involvement. The question is no longer whether AI can produce research-level mathematics, but how institutions, researchers, and AI companies will adapt to a world where it can.