OpenAI publishes a proof of Navier-Stokes, its genesis sparks controversy

OpenAI claims to have solved the Navier-Stokes problem using 10,000 agents. The proof is public, but its origin and attribution are subject to debate.

Roughly 10,000 agents, 88 hours of research, and 130 billion output tokens were used to tackle one of the most difficult problems in mathematics. On September 8, 2026, OpenAI published a proof claiming that the Navier-Stokes equations can develop a singularity in finite time. The announcement could have remained a landmark scientific moment. Instead, it immediately triggered a dispute over the origin of the ideas, the possible use of confidential work, and the way the company allegedly tried to allocate credit.

The Navier-Stokes equations describe fluid motion, from air flowing around a wing to ocean currents. They combine velocity, pressure, viscosity, and external forces to represent a fluid as a continuous medium.

Their routine use in engineering does not mean their mathematical behavior is fully understood. In three dimensions, no one had proved whether an initially smooth solution must remain smooth indefinitely or whether certain quantities could become infinite after a finite period.

Such a breakdown is called a singularity, or “blowup.” It does not mean that a real fluid would suddenly begin moving at infinite speed. It instead shows that, under certain conditions, the mathematical model ceases to provide the expected smooth solution.

As early as 1934, Jean Leray established the existence of so-called weak solutions, which satisfy the equations in a broader sense. The questions of their regularity and uniqueness nevertheless remained open. In 2000, the Clay Mathematics Institute included the puzzle among its seven Millennium Prize Problems, each carrying a $1 million award.

The proof published by OpenAI runs for 166 pages. For any positive viscosity, it constructs a three-dimensional flow that starts at rest and is driven by an external force that is smooth and confined in space and time. Its velocity becomes unbounded at the terminal time, while its total kinetic energy remains bounded.

The presence of an external force quickly caused confusion in online discussions. The best-known version of the problem is often summarized as asking whether a fluid with no external force can produce a singularity. OpenAI’s document does not establish that scenario.

It does not, however, address a question outside the scope of the prize. The official formulation written by Charles Fefferman provides four possible routes. Options A and B involve proving the global existence of smooth solutions without forcing, either in ordinary three-dimensional space or in a periodic domain. Options C and D allow a counterexample with an external force, provided that the force meets specific smoothness conditions.

OpenAI says it establishes C in three-dimensional space and D on a periodic torus. If every argument is correct and the formalized assumptions precisely match Fefferman’s requirements, the proof would therefore answer one of the accepted versions of the Millennium Problem.

The construction takes the form of a vortex that contracts inward. The fluid spirals toward the center as its core becomes increasingly narrow and elongated. Angular and axial velocities increase, but the affected region shrinks quickly enough to keep the total energy from diverging.

The real obstacle is preserving a perfectly smooth external force. It would be easy to define a singular force after the fact that produces the desired behavior, but that would not answer the problem. The proof therefore introduces oscillatory pulses around the core. Their momentum transfers correct the divergent terms produced by the main vortex.

A sequence of additional corrections must then cancel the remaining errors at increasingly fine scales. The residual force stays smooth until the moment when velocity becomes unbounded. This cascade of compensations accounts for much of the document’s length and technical complexity.

The result does not directly provide a new way to simulate weather, design an aircraft, or predict turbulence. It concerns the existence of extreme behavior within an idealized mathematical framework. Its importance lies in what it reveals about the internal limits of the equations, rather than in any immediate industrial application.

The way the proof was produced is almost as unusual as its content. In its official presentation, OpenAI says it used a next-generation internal model whose training began on August 28. The company describes it as significantly more capable than GPT-6 Astra, but has not released its name, architecture, detailed results, or expected availability date.

On September 1, OpenAI’s teams reportedly heard rumors that two Millennium Prize Problems had been solved. They then assigned several groups of agents to the six remaining open problems and to a set of questions considered more accessible.

Each group received a different formulation. Some attempted to prove that Navier-Stokes solutions remain smooth, while others tried to construct a breakdown. The agents could consult a cached copy of the internet, run code, communicate with one another, and share intermediate findings.

Nearly 100 agents reportedly spent about 50 hours working first on the Euler equations, which describe a fluid without viscosity. OpenAI says this first group produced another singularity proof, this time without an external force.

The company then redirected its resources toward Navier-Stokes. Codex was used to consolidate ideas considered useful across the different groups and feed them into new instructions. The system assigned to the main problem eventually reached around 10,000 concurrent agents.

OpenAI reports 2.7 million messages and 130 billion output tokens for this research alone. Across all the problems attempted, the agents reportedly exchanged 4.9 million messages and generated about 300 billion tokens. Several company executives estimated the computing cost at several million dollars.

The agents reportedly reached the result on September 5, 88 hours after the experiment began. GPT-6 Astra then spent another 17 hours formalizing and checking