Summary

OpenAI says an internal multi-agent AI system produced an analytical proof of finite-time singularity formation in a three-dimensional Navier–Stokes flow and formalized it in Lean. The company is presenting the work as a possible resolution of the Millennium Prize problem while saying it will not claim the prize.

Contents

What OpenAI is claiming

OpenAI published its account on September 8, 2026, under the title “On the Navier–Stokes Millennium Prize Problem”. The company says an internal AI system produced an analytical proof showing that an initially smooth, stationary three-dimensional incompressible fluid can develop a singularity in finite time under a smooth applied force, while its total energy remains finite.

The claimed result concerns statements C and D in the official Millennium Prize formulation. OpenAI describes the construction as a vortex that spirals inward and becomes increasingly elongated. Its central region shrinks while the local fluid speed increases without bound.

The Navier–Stokes equations apply Newton’s second law to fluid motion while representing the fluid as a continuous medium rather than tracking individual molecules. The existence and smoothness problem asks whether a smooth, constant-density, incompressible three-dimensional flow remains smooth for all time, or whether it can develop a finite-time singularity.

In this setting, a singularity means that fluid speed grows without bound within finite time despite viscosity, the physical effect that normally smooths out differences in motion. OpenAI says its construction keeps the total energy finite even as the velocity becomes unbounded in a shrinking region.

The company also says the result has been formalized in the Lean proof assistant and that Lean verification took an additional 17 hours using GPT-6 Astra.

Why the singularity matters

The claimed mechanism is a form of self-generated fluid breakdown. OpenAI describes the acceleration, pressure-gradient, momentum-transfer and viscosity terms in the equations as becoming large while cancelling in a precise way. That balance allows the vortex to intensify locally without the system's total energy diverging.

This distinction between local velocity and total energy is central. The claim is not that an ordinary physical fluid was observed moving at infinite speed. It concerns whether the mathematical equations admit a solution whose local behaviour becomes singular while an overall energy measure remains finite.

The problem’s modern history includes Leray’s 1934 generalized-existence result, while the question of smoothness remained unresolved. The problem is one of the seven Millennium Prize Problems identified by the Clay Mathematics Institute.

If the proof survives expert examination, it would settle a major question in mathematical fluid dynamics. The equations are used in fields including aircraft design, weather forecasting and blood-flow studies, although the claimed result concerns an idealised mathematical model rather than every real fluid or engineering flow.

What Lean adds

Lean is a formal proof system in which mathematical statements and proof steps are encoded so that a computer can check them according to the system’s formal rules. A Lean formalization therefore provides a machine-checkable representation of the reasoning, rather than only a conventional written argument.

For this result, the formalization is important because it creates a separate object that mathematicians can inspect and run through Lean’s checking process. The interpretation still depends on the theorem that was encoded: researchers need to examine how the formal statements correspond to the intended Navier–Stokes problem and whether the assumptions describe the relevant case.

OpenAI says it is sharing both a written proof and the Lean formalization. The evidence available in the company’s publication is therefore an OpenAI-reported mathematical result accompanied by formal proof materials. Independent mathematical assessment and external acceptance are the next significant stages of evaluation.

OpenAI also says it is not seeking to claim the Millennium Prize for the result.

How the AI effort was organised

OpenAI describes the work as a coordinating multi-agent system rather than a single model producing one uninterrupted solution. Approximately 10,000 concurrent agents were involved in the group that produced the Navier–Stokes result.

According to the company, the agents reached the resolution on September 5, about 88 hours after the first agents were launched. The Navier–Stokes effort generated approximately 2.7 million agent messages and about 130 billion output tokens. Lean formalization and verification required a further 17 hours.

OpenAI says the discovery used an internal model significantly more capable than GPT-6 Astra, but it does not identify that model. GPT-6 Astra was later used for the Lean formalization and verification.

The project also included preliminary work on an unforced Euler-equation regularity question. OpenAI says nearly 100 agents worked on that problem for approximately 50 hours before the Navier–Stokes effort.

The company’s account addresses concurrent mathematical work as well. OpenAI says an investigation found that prompts from Tristan Buckmaster could not have influenced its result, including through training, and that the OpenAI and Anthropic-related proofs differed. Those statements are part of OpenAI’s own account of the project.

How to interpret the result

The immediate significance is the combination of an ambitious mathematical claim with a formal verification artifact. The written proof explains the construction in conventional mathematical language, while the Lean version offers a computer-checkable encoding of the argument.

The central result remains an OpenAI-reported claim pending scrutiny by independent mathematicians. That scrutiny will need to examine the exact problem formulation, the assumptions behind statements C and D, the proof’s mathematical steps and the relationship between the Lean theorem and the official Millennium Prize question.

The work is also a capability demonstration for AI-assisted research. Its broader importance will depend partly on whether other researchers can examine and reproduce the formalized result, and whether future systems can produce similarly verifiable advances on difficult mathematical problems.

Sources