The Pulse
OpenAI Claims Its AI Found a Navier–Stokes Blowup Solution
OpenAI says an internal AI system produced a proof showing that three-dimensional Navier–Stokes flow can develop a finite-time singularity. The claim has drawn scrutiny over mathematical verification, research credit and the use of 10,000 c
AI.info Team ·
“I thought there must be a mistake somewhere. And on Sunday morning we had the final solution, Lean-formalized and everything.”
Sebastien Bubeck, mathematician and AI researcher, OpenAI
OpenAI says an internal artificial intelligence system has produced a solution to the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems that have challenged mathematicians for decades. The company published a 166-page technical paper and a machine-checkable formalization in Lean on September 8, claiming that a smooth three-dimensional fluid can develop unbounded velocity in finite time while its kinetic energy remains bounded.
The announcement describes a mathematical result, not a new way to calculate ordinary fluid flows. Engineers already use numerical approximations of the Navier–Stokes equations for aircraft design, weather models and blood-flow studies. The unresolved question is whether the equations can break down under the conditions specified by the Clay Mathematics Institute, which set the problem in 2000 and attached a $1 million prize to a valid solution.
OpenAI says it will not claim the prize. Its paper and formal proof are now available for mathematicians to examine, and the company’s wording is significant: the result is presented as an AI-generated solution and as evidence of progress in automated research, but the formal status of the Millennium problem still depends on expert review.
OpenAI’s Result Targets Alternatives C and D
The Navier–Stokes equations describe incompressible fluid motion through velocity, pressure, viscosity and external force. In three dimensions, mathematicians know that finite-energy weak solutions exist, but they have not established whether smooth solutions remain smooth for all time. A singularity would mean that the velocity becomes unbounded at a finite time, causing the classical description to fail.
OpenAI’s paper constructs a solution that begins from rest and develops a singularity while retaining bounded kinetic energy. The construction uses a smooth force that is compactly supported in space and time. According to the paper, the velocity remains smooth before the singular time, then grows without bound near a concentrating point.
The proof addresses alternatives “C” and “D” in Charles Fefferman’s official Clay problem statement. Alternative C concerns breakdown on three-dimensional Euclidean space, while alternative D concerns a periodic three-dimensional domain. OpenAI’s public repository describes the two results as showing that smooth initial data and forcing can prevent the existence of a global smooth solution with uniformly bounded kinetic energy in the whole-space case, and can prevent a global smooth solution on the periodic torus.
The construction is built around a collapsing vortex. OpenAI describes a central region that spirals inward, stretches along its axis and becomes increasingly concentrated. The velocity rises as the core shrinks, but the total kinetic energy stays controlled because the region carrying the fastest flow becomes smaller.
That balance is the mathematical difficulty. The acceleration, pressure, nonlinear momentum transfer and viscous terms can each grow large near the singularity. The proof arranges for them to cancel in a way that leaves a smooth external force. Oscillatory pulses in an annular region transport momentum through the fluid and supply the stress needed to sustain the collapsing structure.
10,000 Agents and 88 Hours
OpenAI says the work began after the company heard rumors on September 1 that researchers might have made progress on two Millennium Prize Problems. The company had been training a new internal model since August 28 and decided to test it across the open problems, along with several other difficult mathematical questions.
The system did not consist of one model producing a single answer in one sitting. OpenAI divided the work among groups of coordinating agents, each exploring different versions and approaches. The company says the group that produced the Navier–Stokes result involved roughly 10,000 concurrent agents. The agents could read from a cached version of the internet, run code and communicate within their assigned groups.
OpenAI says the first agents were launched on September 1 and the system reached its Navier–Stokes resolution on September 5, about 88 hours later. A further 17 hours of work went into formalization and verification through GPT-6 Astra, the company’s announced model for checking the proof in Lean.
The scale of the computation is unusual even by the standards of automated mathematics. Across all of the attempted problems, OpenAI says the agents sent 4.9 million messages and generated about 300 billion output tokens. The Navier–Stokes effort accounted for 2.7 million messages and approximately 130 billion output tokens.
OpenAI’s account also shows how the system moved between problems. Agents first produced a result for a related regularity question involving the Euler equations, which omit the viscosity term from Navier–Stokes. The company then shifted agents toward Navier–Stokes and used Codex to consolidate useful ideas from separate agent groups.
Lean Checks the Formal Proof, Not the Whole Claim
Lean provides a formal language in which mathematical statements and their proof steps can be checked by software. OpenAI has released Lean files and a public repository containing certificates associated with the Navier–Stokes and Euler results. That gives outside researchers a way to inspect whether the formalized steps follow within the chosen definitions and axioms.
Formal verification matters because language models can produce persuasive but invalid arguments. A proof assistant can catch many local errors, missing cases and unjustified transformations. It does not, by itself, decide whether the formal statement captures the intended mathematical problem, whether the surrounding exposition is clear or whether a construction has been transcribed correctly from the underlying research.
OpenAI’s paper makes a specific claim: for every positive viscosity, it constructs a smooth force and a solution that begins from zero velocity, remains finite in kinetic energy and develops unbounded velocity in finite time. The proof is far more specialized than a general prediction that every physical fluid will form an infinite-speed vortex. Its conclusion applies to the mathematical conditions set out in the paper.
Independent mathematicians still need to read the argument in full. The Clay Mathematics Institute’s problem statement requires smooth, physically reasonable solutions and distinguishes between whole-space and periodic formulations. Any decision on a prize would require the result to survive scrutiny by experts in partial differential equations and fluid dynamics, not simply pass a software check.
The Announcement Collided With Outside Research
OpenAI’s announcement arrived alongside a dispute involving Tristan Buckmaster, a mathematics professor at New York University, and Levent Alpöge, an Anthropic researcher. Buckmaster said the pair had been working for about a year on related fluid equations with help from several AI models, including OpenAI’s Codex and Anthropic’s Claude.
According to Buckmaster’s account, the researchers made major progress on August 15 and verified it by August 22. Their work focused on a forced Euler problem, a related equation that removes viscosity from Navier–Stokes. Buckmaster said the pair’s approach drew on earlier work by Diego Córdoba and Luis Martínez-Zoroa and was unusual enough that he did not believe another group would arrive at it simply by prompting a model with the problem statement.
OpenAI said it learned of rumors about outside work before launching its own effort. The company has denied using the pair’s prompts or proof to direct its agents. In the September 8 briefing described by WIRED, Bubeck said the researchers and agents had not seen the outside work before it became public.
Buckmaster has not accused OpenAI of stealing the work. He has said he does not know whether his data was used and has questioned whether OpenAI’s systems could have encountered material from his Codex sessions through training or product data. The distinction matters: evidence that two projects followed related mathematical paths would not establish that one copied the other.
The timing has nevertheless placed research credit at the center of the story. OpenAI’s result and the outside researchers’ Euler work are not identical. OpenAI says its proof handles the viscous Navier–Stokes equations without external forcing in the singular mechanism, while the outside work addressed a forced Euler result. Whether the approaches share a deeper mathematical dependency is a question for specialists who can compare the complete proofs.
A Test of AI Research Claims
OpenAI’s announcement puts two standards of success side by side. One is computational: the system generated a long proof, formalized it and placed the underlying files online. The other is scientific: mathematicians must determine whether the construction is correct, whether it satisfies the exact Clay formulation and what ideas can be understood and reused by people.
The distinction is important because a finished proof can conceal the path that produced it. Human mathematicians often learn from failed approaches, intermediate lemmas and the reasons a particular construction works. A multiagent system can search many directions at once, but its speed may leave researchers with a result that is difficult to interpret even when the formal checker accepts it.
OpenAI frames the work as a report on the capability of an internal model that is more powerful than GPT-6 Astra, rather than as a prize claim. That framing avoids declaring the Millennium problem settled by the company alone, but it also places pressure on the published paper and repository. The technical record will have to carry the claim beyond OpenAI’s own description.
For now, the verified fact is narrower than the headline that will travel around the internet: OpenAI has released a proof claiming finite-time blowup for a specified class of three-dimensional Navier–Stokes solutions, along with a Lean formalization and source files for inspection. The Clay prize remains unclaimed, and the mathematical verdict belongs to the researchers who examine the proof line by line.