OpenAI says an experimental artificial intelligence system has produced a solution to the Navier-Stokes Millennium Prize Problem, one of mathematics’ most difficult unsolved questions, after roughly 10,000 AI agents worked in parallel for 88 hours.

The result, announced Sept. 8, could become a landmark in both mathematics and artificial intelligence. If the proof survives independent scrutiny, it would resolve a problem that has resisted generations of mathematicians for roughly 90 years. It would also provide one of the strongest demonstrations yet that AI systems can move beyond assisting researchers and contribute directly to original mathematical discovery.

But understanding what OpenAI did requires looking beyond the headline.

The Navier-Stokes equations describe how fluids move. Water flowing through a pipe, air moving around an aircraft wing, blood traveling through arteries and large-scale atmospheric flows can all be studied using equations from this family.

The problem is that mathematicians have never fully understood what happens to the equations under every possible set of conditions.

A fluid may begin with perfectly smooth motion. The unresolved question was whether its mathematical description must remain smooth forever, or whether the equations can reach a point where velocity grows without limit in a finite amount of time. Mathematicians call such a breakdown a singularity.

OpenAI says its system demonstrated that this breakdown can occur.

Its proof constructs an initially smooth three-dimensional fluid that starts at rest. A smooth external force is applied, creating a vortex that becomes increasingly stretched while spiraling inward. According to OpenAI, the process eventually causes velocity to become unbounded in finite time even though the fluid’s total energy remains finite.

That satisfies statements “C” and “D” in the Clay Mathematics Institute’s official formulation of the Millennium Prize Problem, OpenAI says.

The result is mathematically striking because viscosity normally works against this kind of behavior. Viscosity is the internal friction that helps smooth out the motion of fluids. Honey has relatively high viscosity, while water has much less. Researchers had to determine whether that smoothing effect could always prevent a singularity from forming.

OpenAI’s proposed proof says it cannot.

An AI Research Team With Thousands of Agents

The scale of the computation is almost as significant as the mathematical result.

OpenAI did not simply give the problem to one chatbot and wait for an answer. The company organized large groups of autonomous AI agents powered by an unreleased internal model that it says is significantly more capable than GPT-6 Astra.

The agents could run code, consult a cached version of the internet and communicate with other agents within their assigned groups. Different groups explored different approaches. OpenAI then used Codex to consolidate promising ideas and feed useful discoveries back into further investigations.

OpenAI describes the successful group as involving “on the order of 10,000 concurrent agents.” That distinction matters. The company has not released an exact integer count, so reports stating precisely 10,000 should be understood as an approximation.

The scale of their communication was extraordinary.

For the Navier-Stokes effort alone, OpenAI says the agents exchanged 2.7 million messages and generated approximately 130 billion output tokens.

The agents reached the mathematical result after about 88 hours, or three days and 16 hours. OpenAI then used GPT-6 Astra for another 17 hours to formalize and verify the proof in the Lean theorem prover.

That puts the process at roughly 105 hours, or four days and nine hours, from the start of the agent effort through formal verification.

There is no scientifically valid way to translate that directly into the number of years a human mathematician would require. The problem itself demonstrates why: mathematicians had been working around the underlying question for decades without a complete solution.

The token count does, however, illustrate the enormous amount of parallel work involved.

OpenAI says one English token corresponds roughly to three-quarters of a word. Using that rough conversion, 130 billion tokens would correspond to about 97.5 billion words. A person typing continuously at 50 words per minute would need roughly 3,700 years simply to produce that amount of text.

That comparison is only about output volume. AI tokens are not equivalent to human reasoning, and much of the agents’ output would have included failed approaches, intermediate arguments, calculations and communication between agents. Still, it gives a sense of the computational scale behind the 88-hour result.

Why This Matters for Mathematics

For mathematicians, the significance extends beyond Navier-Stokes.

AI systems have previously helped with proofs, searched enormous mathematical spaces and resolved narrower open questions. But Quanta Magazine described the Navier-Stokes result, if it withstands scrutiny, as potentially the most important mathematical proof yet produced by an AI system.

It also suggests a different model for doing mathematics.

Instead of one researcher or a small team pursuing a few promising ideas, thousands of AI agents can explore competing approaches simultaneously. Failed paths can be discarded, promising arguments shared across groups and successful ideas formally checked by another system.

That could change the economics and speed of mathematical research, particularly on problems where the space of possible approaches is too large for any individual researcher to investigate.

Formal verification adds another layer. OpenAI released a version of the proof in Lean, a system that checks whether each mathematical step follows from precisely stated assumptions.

That does not mean the Millennium Prize has already been awarded.

The Clay Mathematics Institute requires a proposed solution to be published in a qualifying outlet, remain published for at least two years and achieve general acceptance within the global mathematics community before the institute will consider awarding its $1 million prize. OpenAI has also said it does not intend to claim the prize.

Independent mathematicians are now examining the work. Martin Bridson, president of the Clay Mathematics Institute, described the announcement to Nature as an “exciting day,” while Fields Medal-winning mathematician Terence Tao stressed the importance of conventional professional scrutiny and publication.

Does It Change the Practical World?

Not immediately.

Engineers already use Navier-Stokes-based models every day in aircraft design, weather research, fluid engineering and studies of blood flow. Those applications do not suddenly stop working because a mathematically constructed singularity exists.

There is another important limitation. Real fluids are made of molecules and atoms, while the Navier-Stokes equations treat fluid as a continuous substance that can theoretically be examined at endlessly smaller scales.

The singularity constructed in the proof eventually reaches behavior that cannot physically occur: infinite velocity. Quanta notes that this means the result has no immediate engineering consequence for real fluids.

Its deeper importance is foundational.

The result would show that the Navier-Stokes equations have a genuine mathematical limit. Under certain carefully constructed conditions, the continuum model can break down even when everything begins smoothly.

That knowledge can influence future work on turbulence, fluid simulation, numerical methods and the mathematical assumptions behind physical models. It tells researchers more precisely where one of science’s most important sets of equations can and cannot be trusted as an idealized description of reality.

And the larger impact may come from how the answer was found.

A problem that occupied human mathematics for nearly a century was attacked by thousands of cooperating artificial researchers, reduced to a proposed solution in 88 hours and formally verified 17 hours later.

Featured image via X screengrab.