The cleanest version of the story is also the most precise one.

On September 8, OpenAI published a construction in which an initially smooth three-dimensional incompressible Navier-Stokes flow develops a singularity in finite time. The fluid begins at rest, is subjected to a smooth external force and retains finite energy as its velocity becomes unbounded.

OpenAI says this establishes alternatives C and D in the official Millennium Prize formulation.

On September 11, the Clay Mathematics Institute responded by saying the Navier-Stokes problem had apparently been settled. It has not yet awarded a prize or issued a final judgment on attribution.

The proposed singularity is a vortex collapsing inward

Navier-Stokes equations are used to describe fluid motion, from air around aircraft to weather systems and blood flow.

The mathematical problem asks whether smooth three-dimensional solutions can always remain smooth or whether the equations can produce a singularity in finite time.

OpenAI's construction follows a vortex that spirals inward while stretching along its axis. The core becomes smaller as the circulating velocity grows without bound.

The difficult part is not merely making quantities large. The applied force itself must remain smooth. Acceleration, pressure gradients, momentum transport and viscosity have to grow and cancel with extraordinary precision while the solution approaches its singularity.

The forced and unforced questions are not the same

This is the technical distinction most likely to disappear in headlines.

OpenAI's construction contains smooth external forcing. It does not show that an unforced viscous three-dimensional fluid spontaneously blows up from smooth initial data.

That does not automatically exclude the result from the Millennium Prize Problem. The official Clay formulation contains four alternatives, labelled A through D, and the forced counterexample alternatives C and D are explicitly part of it.

The unforced global-regularity question remains distinct and important.

So two statements can coexist: OpenAI's claimed result may settle the official Millennium Prize formulation while the famous unforced version remains open.

This was not one model sitting on one prompt for 88 hours

OpenAI's research process is arguably the second major story.

Beginning September 1, the company deployed groups of coordinating agents with access to a cached version of the internet and code-execution tools. Separate groups were given different variants of the open Millennium Prize problems and encouraged to explore different strategies.

The Navier-Stokes effort eventually involved on the order of 10,000 concurrent agents.

OpenAI says they reached the announced construction after about 88 hours. Lean formalization and verification then took another 17 hours using GPT-6 Astra.

The Navier-Stokes portion alone generated about 2.7 million inter-agent messages and approximately 130 billion output tokens.

Euler came first

The path to the result did not begin with Navier-Stokes alone.

OpenAI also assigned agents to easier related questions. Nearly 100 agents spent about 50 hours on a regularity problem for the Euler equations, which remove the viscosity term from Navier-Stokes.

After producing an unforced Euler blowup construction, OpenAI redirected more resources toward Navier-Stokes and fed useful intermediate ideas across agent groups.

Codex was used to consolidate promising insights before they were redistributed.

That makes the system closer to a large parallel research organization than a conventional question-answering model.

Lean makes the artifact checkable, not instantly accepted

OpenAI released both a conventional mathematical write-up and a Lean formalization.

Formal proof assistants can verify that a chain of deductions follows from its stated definitions and axioms. That removes a large class of mistakes that can hide inside long handwritten arguments.

It does not eliminate the broader academic process.

Mathematicians still need to inspect whether the formal statement precisely matches the claimed theorem, understand the construction and determine how priority and credit should be assigned.

The Clay Mathematics Institute has explicitly said that its evaluation process will be deliberately unhurried.

OpenAI says it does not intend to claim the $1 million prize.

A priority dispute arrived almost as quickly as the proof

Another research group was working on closely related problems at the same time.

NYU mathematician Tristan Buckmaster and Anthropic employee Levent Alpöge had been using AI tools in their own work on fluid-equation singularities. OpenAI says rumors about two Millennium Prize breakthroughs helped motivate its September 1 push.

The company later learned that the rumor involved Buckmaster and Alpöge's Euler work rather than an identical completed Navier-Stokes solution.

That timeline produced a public fight over research priority, attribution and whether private interactions with AI tools could somehow have influenced the result.

OpenAI says an internal investigation found that Buckmaster's Codex prompts from the preceding two months could not have influenced the model, including through training. Disagreement over how the broader race and credit discussions were handled has not disappeared.

Navier-Stokes may not be the biggest AI implication

Millennium Prize Problems carry enormous symbolic weight. Solving one is an unmistakable marker of mathematical importance.

But the more transferable development may be the method.

OpenAI has demonstrated a system that can throw thousands of agents at an unsolved scientific question, explore competing paths, merge intermediate discoveries and end with a machine-checkable mathematical artifact.

It is enormously compute-intensive and nowhere near an everyday scientific tool.

Still, this is qualitatively different from asking a model to reproduce mathematics whose answer is already known. The system was deployed as part of a research process intended to create a new result.

Clay has not yet removed every qualification from the verdict.

For now, “apparently settled” is remarkable enough.