The claim is a finite-time singularity

Navier-Stokes equations model fluid motion and are used across aerodynamics, weather modelling, engineering and biological flows. The Millennium Prize question asks something much more fundamental than whether engineers can use the equations successfully.

For three-dimensional incompressible flow, can perfectly smooth initial conditions remain smooth forever, or can the equations produce a singularity in finite time?

OpenAI says its proof establishes the latter.

Its construction begins with a smooth fluid at rest under a smooth external force. A vortex then spirals inward and stretches along its axis while a central region contracts and its velocity grows without bound. The total energy remains finite, according to the proof.

OpenAI argues that this establishes statements C and D in the official Millennium Prize formulation.

This was computational research at industrial scale

OpenAI says it began the broader experiment on September 1 after hearing rumors that major mathematical problems had recently been resolved elsewhere.

The company assigned several open problems to groups of agents powered by an internal model that it describes as significantly more capable than GPT-6 Astra. That model has been training since August 28 and was still being improved during the experiment.

Agents could run code, read from a cached version of the internet and communicate within coordinated groups. OpenAI says the Navier-Stokes effort involved on the order of 10,000 concurrent agents.

They reached the selected solution on September 5, about 88 hours after the first agents were launched.

For Navier-Stokes alone, those systems exchanged roughly 2.7 million messages and generated approximately 130 billion output tokens.

A Lean proof is a substantial difference

Language models are capable of producing mathematical arguments that look persuasive while containing subtle gaps. OpenAI did not stop with a natural-language manuscript.

The company also released a formal proof in Lean, where each encoded step is checked mechanically by a proof assistant. OpenAI says GPT-6 Astra completed formalization and verification in another 17 hours.

That raises the evidentiary bar considerably.

It does not make independent mathematical review unnecessary. Researchers still need to examine whether the formal statements faithfully encode all the assumptions and requirements of the Millennium Prize formulation, as well as understand the construction itself.

The authorship dispute arrived almost immediately

The timing has also created an uncomfortable research-data controversy.

NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge had been pursuing closely related work on the Euler equations. Their research workflow included OpenAI Codex as well as Anthropic tools.

Buckmaster has publicly questioned whether information from private Codex work could have influenced OpenAI's internal model or research direction.

OpenAI says no specific user data from the researchers was accessed to solve the problem. It also says it cannot completely rule out the possibility that de-identified data derived from their product usage contributed to model improvement.

Those are materially different claims. There is currently no demonstrated evidence that OpenAI copied their unpublished proof into its Navier-Stokes result. The episode still exposes a difficult problem for AI-assisted science: researchers may increasingly conduct confidential work inside tools operated by companies capable of becoming research competitors themselves.

The million-dollar prize is not the immediate issue

The Clay Mathematics Institute assigns a $1 million prize to each Millennium Prize Problem. OpenAI says it does not intend to claim the award.

Prize money is secondary here. Mathematical acceptance requires time, external scrutiny and a much broader understanding of the proof than a company announcement can provide one day after publication.

There is another result hidden inside the announcement, though. OpenAI says the model responsible for the discovery is still being trained and already exceeds GPT-6 Astra substantially in mathematics.

If the proof survives, one historic story will be about a fluid equation. The other will be about what kind of research system needed only 88 hours to attack it.