A question that sat open for roughly 90 years was answered in 88 hours — about 9,000 hours of unsolved mathematics retired for every hour a machine spent on it. On September 8, 2026, OpenAI published a proof, produced by a swarm of AI agents running on an unreleased internal model, that the Navier–Stokes equations we use to model every fluid from jet airflow to blood can break themselves.

Before that means anything, you need to know what was actually broken. So let's start there.

First: what problem did OpenAI solve?

The Navier–Stokes equations are Newton's F = ma rewritten for fluids. They date to the nineteenth-century work of Claude-Louis Navier and George Gabriel Stokes, and they are the workhorse behind aircraft design, weather forecasting, and the study of blood flow. Every time a simulation tells an engineer where turbulence will form on a wing, these equations are underneath it.

They contain one enormous simplification: they treat a fluid as a continuous medium. Not trillions of jostling molecules — a smooth, infinitely divisible substance you can zoom into forever. That approximation is why the equations are usable at all. It is also the crack in the foundation.

The open question, in plain terms: if you start a fluid moving smoothly, can the equations themselves make it go haywire? Formally — can a three-dimensional incompressible fluid of constant density develop a singularity, a point where speed shoots to infinity in a finite amount of time, even though viscosity, the fluid's internal friction, is constantly trying to smooth everything out?

That matters because a real fluid cannot move infinitely fast. If the equations produce infinite speed, they have stopped describing reality, and you would have to go back to tracking individual particles. So the real question is: do our fluid equations have a self-destruct button, and where is it?

In 1934, Jean Leray proved that solutions exist in a weaker, generalised sense — but whether they always stay smooth was left hanging. In 2000, the Clay Mathematics Institute named Navier–Stokes existence and smoothness one of seven Millennium Prize Problems, each carrying a $1 million award. Only one has ever been settled: the Poincaré conjecture, proved by Grigori Perelman, who was awarded the prize in 2010 and declined it.

There is a catch in how a result like this gets officially accepted, and it is why OpenAI is not asking for the money. More on that near the end.

88 h
From launch to proof
~10,000
Concurrent agents on it
130B
Output tokens spent
~90 yrs
The question stood open

What the proof says: a fluid that tears itself apart

OpenAI's system produced an analytical write-up and a formalization in Lean showing that a fluid which starts smooth and at rest can develop a singularity in finite time. In the Clay problem's official formulation, that establishes statement "C" (and also "D") — the disproof side of the question. Smoothness does not always survive.

The object at the centre of it is a vortex: a spinning swirl that spirals inward while stretching out lengthwise, which OpenAI describes as becoming like spaghetti. The core shrinks and speeds up at once, and the balance between those two motions lets velocity run to infinity while the fluid's total energy stays finite, exactly as physics requires.

That constraint is the whole difficulty. Anyone can break an equation by shoving infinite energy into it; the bar here was that the breakdown had to come from the fluid's own motion, with the applied force staying smooth and the energy bounded throughout. Less like smashing a glass, more like shattering one with a sung note.

The eight-day sprint

Timeline of OpenAI's Navier–Stokes agent run, from model training to publication

From a rumour on September 1 to a Lean-verified proof on September 6. (Chart: BougainWell · Data: OpenAI)

Since August 28 OpenAI had been training a new internal model — still training, still improving — that it says is significantly more capable than GPT‑6 Astra. On Tuesday, September 1, the company heard rumours that two Millennium Prize problems had been resolved, and pointed the model at every open Millennium problem at once. Agents ran in communicating groups with a cached copy of the internet and the ability to execute code; the group behind the Navier–Stokes result ran on the order of 10,000 concurrent agents.

They warmed up on easier neighbours first, including the same blowup question for the Euler equations — Navier–Stokes with viscosity switched off. Roughly 100 agents working for about 50 hours cracked the unforced version, and that surprise redirected everything: OpenAI pulled agents off the other problems, seeded them with the Euler resolution, upgraded them mid-run when a further-trained model landed, and used Codex to carry insights from one group into another. The proof arrived on Saturday, September 5, about 88 hours after launch; Lean formalization and verification took another 17 hours via GPT‑6 Astra.

The rumour that started it all, OpenAI notes, traced back to Levent Alpöge and Tristan Buckmaster, who had resolved the forced Euler problem — and OpenAI says it recognises their priority on that result.

What it cost, in machine terms

Share of OpenAI's total agent run that went to the Navier–Stokes problem

More than half the messages and over 40% of the tokens in the entire multi-problem run went to one equation. (Chart: BougainWell · Data: OpenAI)

The agents sent 4.9 million messages and generated about 300 billion output tokens across everything they attempted; Navier–Stokes alone took 2.7 million messages and roughly 130 billion tokens. Run the division: about 61,000 output tokens for every message actually sent. Almost everything these agents produced went nowhere, discarded on the way to the fraction that survived. That is the real shape of automated research — not a machine writing a proof, but a machine writing an ocean of wrong proofs fast enough that a right one falls out.

Why the Lean file matters more than the announcement

A model producing a hundred pages of confident-sounding proof is worth very little on its own — plausible mathematical prose is exactly the failure mode you would expect. A Lean file that compiles is a different kind of object: a proof assistant mechanically checks every logical step, and it does not care how the argument was found. Verification took 17 hours, about 16% of the total time, and it is that 16% which makes the other 84% checkable at all.

And notice what the winning recipe was: not one model thinking harder, but thousands of agents attacking many framings of a problem in parallel, cross-pollinated by another model harvesting whatever was working. That generalises anywhere a problem has a cheap correctness check — a compiler, a test suite, a simulator, a proof assistant. Generate wastefully, verify ruthlessly, keep the survivors. The bottleneck is never generation; it is whether you have an oracle that can tell you when you are right.

Not yet a settled result. A published proof and a Lean formalization are strong evidence, but the mathematical community still has to read and probe the write-up, the exact statement it addresses, and the formalization's assumptions. Treat this as a claim with unusually good receipts, not a closed case.

The catch: nobody is collecting $1 million

Here is the open loop, paid off. Clay does not hand over a prize for a preprint. Its rules require publication in a refereed mathematics journal of worldwide repute, at least two years elapsed after that publication, and general acceptance in the global mathematics community — the waiting period exists precisely so errors have time to surface.

OpenAI has said plainly it does not intend to claim the prize, framing the release as a report on the pace of its models. Take that at face value, but notice the incentives too: claiming a Millennium Prize invites a two-year adversarial review of your proof, while announcing a capability invites a re-rating of your company. Only one of those pays out immediately.

What to Take Away

  • The Navier–Stokes problem asked whether smooth fluid motion can blow up on its own. OpenAI's answer is yes: a smooth fluid at rest, under a smooth force and with finite energy throughout, can reach infinite speed in finite time. That is a limit on our equations, not a discovery about water.
  • Finite energy was the hard part, not infinite speed. Any equation breaks if you feed it infinite force; making it break from its own internal balance is the achievement.
  • Remember the ratio: ~61,000 output tokens per message sent. Machine research is mostly discarded thinking. Judge these systems by verified output, not fluency.
  • Formal verification is the new gate. A Lean file that compiles is worth more than a hundred pages of confident prose, and that gap widens as generation gets cheaper.
  • A proof is not a prize. Clay requires journal publication, two more years, and community acceptance. Announcements move markets in days; mathematics still moves in years.

Cover image: OpenAI. Charts: BougainWell, built from OpenAI's published account of the Navier–Stokes result. This article is for general information only and is not investment advice.

Sources

All analysis and opinions in this article are BougainWell's own.