OpenAI has unveiled a Navier–Stokes proof that it says resolves the celebrated fluid-dynamics problem using about 10,000 concurrent AI agents. The September 8 release includes a 166-page analytical argument and a machine-checked formalization in Lean, but it has not yet been accepted by the Clay Mathematics Institute.

 

The research has four unusually consequential parts:

  • A claimed finite-time singularity for three-dimensional fluid flow
  • About 10,000 collaborating agents working for 88 hours
  • A Lean formalization completed over another 17 hours
  • A Clay review process that cannot conclude for at least two years

 

OpenAI's Navier–Stokes Proof Targets Finite-Time Blowup

The Navier–Stokes equations describe how fluids such as air and water move. Their three-dimensional behavior has resisted proof for decades: mathematicians did not know whether smooth starting conditions must remain smooth forever or could evolve into a singularity where velocity becomes unbounded.

 

OpenAI's construction takes the second route. Its system produced a vortex that spirals inward and stretches while accelerating, with the central region shrinking fast enough for velocity to diverge even though the flow's total energy remains finite.

 

The company says the example begins with a smooth fluid at rest and applies a smooth external force. That qualification is crucial. Earlier work could generate blowup with rough or artificial forcing, but the official Clay formulation requires the force and initial conditions to satisfy strict smoothness and decay properties.

 

The released argument claims to establish statements C and D in the official problem description. Those statements ask for a counterexample to global smoothness in ordinary three-dimensional space or its periodic counterpart.

 

10,000 AI Agents Used 130 Billion Output Tokens

OpenAI says it began training the internal model on August 28 and launched the mathematics effort on September 1. The model is not publicly available and, according to the company, is significantly more capable than GPT-6 Astra on the relevant reasoning tasks.

 

Groups of agents received different versions of several open problems, with access to cached web material and code execution. Researchers moved promising ideas between groups, and Codex consolidated useful intermediate results so that later agents could work from a broader pool of approaches.

 

The Navier–Stokes run reached its result after roughly 88 hours. It generated 2.7 million inter-agent messages and about 130 billion output tokens, making the proof a product of large-scale parallel search rather than a single conversational response.

 

Before that run, nearly 100 agents spent about 50 hours producing a result for the related Euler equations, which model fluids without viscosity. OpenAI then redirected more compute toward Navier–Stokes and supplied agent groups with the Euler insights that appeared most useful.

 

Related Coverage

 

Lean Verification Narrows, Not Ends, Human Review

OpenAI published both the analytical paper and a public repository containing Lean certificates. Formal verification checks that every encoded logical step follows from stated definitions and assumptions, sharply reducing the risk of an unnoticed algebraic or inferential gap.

 

It does not automatically prove that the encoded theorem is exactly equivalent to the original mathematical question. Human experts must still inspect the translation between the Clay formulation, the analytical construction and the formal statement, as well as assumptions imported into the proof environment.

 

Independent reporting by Quanta Magazine found that the Lean result gives mathematicians substantial confidence, while also documenting unresolved questions about priority and the intellectual path to the proof.

 

Researchers Diego Córdoba and Luis Martínez-Zoroa developed the analytic cascade strategy on which the new work builds. Separately, NYU professor Tristan Buckmaster and Anthropic researcher Levent Alpöge announced AI-assisted advances on related Euler problems, illustrating how several teams converged on similar methods at nearly the same time.

 

Clay Rules Keep the $1 Million Prize Unsettled

The Clay Mathematics Institute still lists Navier–Stokes as an active problem. Its prize rules do not accept direct submissions and require a proposed solution to appear in a qualifying outlet, remain public for at least two years and gain general acceptance from the global mathematics community.

 

That process prevents a release-day headline from becoming an official resolution. Even if specialists find the proof correct, publication venue, scrutiny and community consensus must all be established before the institute can consider awarding the $1 million prize.

 

OpenAI says it does not intend to claim the money. Its stated purpose is to demonstrate the scientific capability of an internal model and the scale at which coordinated agents can explore, combine and formalize technical arguments.

 

The result therefore carries two tests. Mathematicians must determine whether the construction fully satisfies the Navier–Stokes problem, while AI researchers must assess whether millions of agent messages and billions of tokens can become a repeatable method for discovery rather than a one-off computational campaign.

 

If the proof survives both forms of scrutiny, the lasting milestone may be broader than fluid dynamics. It would show that frontier systems can organize thousands of specialized agents around a research program, preserve useful intermediate work and deliver a result precise enough for formal checking and expert review.