Did AI Solve The Navier-Stokes Problem? What Was Proved, And Why It Isn't Over

In September 2026, OpenAI said a team of about 10,000 AI agents had proved that the Navier-Stokes equations, which describe how liquids and gases flow, can break down: a fluid at rest, pushed by a smooth outside force, reaches speeds that grow without limit in a finite time. That matches one of the four statements the $1 million Millennium Prize accepts, and the Clay Mathematics Institute says the problem has “apparently been settled,” pending its own slow review. Some experts call that outside force a loophole, because the question most of them had in mind, whether a fluid left alone can break down, is still open.

Stir a cup of coffee, then lift the spoon out. The swirl spins, slows and dies, every single time. You have never seen a cup of coffee spin itself faster and faster until it tore a hole in the kitchen.

Physics has a set of equations for that swirl, and they date to the early 1800s. They run weather forecasts and shape jet wings. Yet in all that time, nobody could prove the equations agree with your coffee. Could the math ever say that a swirl speeds up without limit? There is a $1 million prize for the answer.

On 8 September 2026, OpenAI announced that its AI had found that answer. Three days later, the keepers of the prize said the problem had “apparently been settled.” And yet the question most experts had in mind is still open, so both things are true at once. To see how, we need a little science first, and then the fine print of the prize.

What Are The Navier-Stokes Equations, In Simple Terms?

Back in the cup, push the spoon and the coffee moves. Now shrink your view to one small blob of coffee. Why does that blob move the way it does?

Three things push on that blob, and two of them come from the coffee around it. That coffee presses on it, which is pressure, and it also drags on it, which is friction. The friction inside a fluid has its own name, viscosity. Honey has a lot of it and water has little. The third push comes from outside the fluid, and it could be gravity or your spoon.

That is all the equations say. In the prize's official problem statement, Charles Fefferman of Princeton writes that the main equation “is just Newton’s law f = ma for a fluid element.” Force equals mass times acceleration, the same law that governs a thrown ball. In words:

acceleration of the blob = friction − pressure push + outside force

And in symbols, as the prize states it:

∂u/∂t + (u · ∇)u = ν∆u − ∇p + f

Here u is the velocity of the fluid, p is its pressure, and ν (the Greek letter nu) is its viscosity. The last letter, f, is the outside force. You do not need to read the symbols, but notice that little f at the end, because the whole story turns on it.

A second, shorter equation adds that the fluid cannot be squeezed into a smaller space. Claude-Louis Navier in France and George Stokes in England worked the equations out separately. They built on the older Euler equations, which describe a fluid with no friction at all.

The same few lines cover the calm stream from a tap and the churning wake of a jet. The Clay Mathematics Institute says they should explain “both the breeze and the turbulence.” That is a lot of work for one law you met in high school.

Has Anyone Ever Solved The Navier-Stokes Equations?

Not in general. NASA tells its students that the equations are “too difficult to solve analytically.” In plain terms, nobody can write down a tidy formula for what a real flow will do next.

So engineers cheat, with style, by chopping the air around a wing into millions of tiny cells. Then a computer works out close-enough answers, one small time step after another. NASA calls this computational fluid dynamics, or CFD. It is the same trick behind numerical analysis in general. OpenAI's post notes that the equations are used for “aircraft design, weather forecasting, and the study of blood flow.” That is how weather forecasting works. It is also how engineers study the turbulence that rattles your flight.

Nobody can solve these equations, and we fly on them anyway.

The $1 million question is more basic than getting numbers out. The Clay Institute asks: “do solutions exist, and are they unique?” Put another way, do the equations always have a sensible answer to give, or can they, in some strange case, give up?

What Does It Mean For A Fluid To “Blow Up”?

Experts call a flow smooth when nothing in it jumps. Speed changes gently from place to place, and it stays finite everywhere. Your coffee is smooth in this sense, even mid-stir.

The opposite is called blow-up, and its formal name is a singularity. OpenAI's post describes it as “speeds in the fluid growing without bound within a finite amount of time.” Picture one point in the cup where the flow gets faster and faster. It passes every speed you can name, and it does so before the clock hits a set moment. Oddly, the total energy in the cup stays finite the whole time.

Real coffee cannot do this. As Quanta Magazine explains, the equations assume you can zoom in on a fluid forever. Zoom far enough, though, and you hit molecules. At that point, OpenAI writes, you would have to track “each particle individually.” So a blow-up is a failure of the model, not of the mug. Your kitchen is safe.

Why Has No One Solved Navier-Stokes?

Because two effects fight inside the equations, and for about 90 years nobody could call the winner.

On one side is friction, the peacemaker. It smooths out sharp changes, which is why your swirl dies. Take friction away and you get Euler's wilder equations.

On the other side is the way a swirl can feed itself. In a flat, two-dimensional flow, friction wins. Fefferman's text says that case has “been known for a long time,” and credits the mathematician O. Ladyzhenskaya. Then he adds a warning. That result “gives no hint about the three-dimensional case, since the main difficulties are absent in two dimensions.”

In three dimensions, a spinning tube of fluid can stretch, thin out and spin faster. Big swirls hand their energy to smaller ones, and that handing-down is the heart of the turbulence behind a gust of wind. The worry was always a runaway. What if the energy keeps rushing into smaller and smaller swirls, faster than friction can calm it?

The best tool experts have is a cap on the fluid's total energy, and it is not enough. The mathematician Terence Tao took up this question in a well-known 2007 essay. Tools like that, he wrote, are “much weaker at controlling fine-scale behaviour” than the large-scale kind. They can see the whole cup, but they go blind right where the trouble would start.

For most of those years, experts bet on friction. “Ten years ago, nobody believed there was a singularity for Navier-Stokes,” the mathematician Diego Córdoba told Quanta.

What Does The $1 Million Navier-Stokes Problem Ask?

In 2000, the Clay Mathematics Institute picked seven famous unsolved problems. It put $1 million on each. Only one had fallen before this year, the Poincaré conjecture, proved by Grigoriy Perelman.

Fefferman wrote the official Navier-Stokes entry. Instead of one question, he offered four statements, and a proof of any one of them wins.

1. Statements A and B say the flow always stays smooth, whatever smooth start you give it. A covers a fluid that fills all of space. B covers a fluid in a box that repeats forever in every direction.
2. Statements C and D say the opposite. Somewhere there is a smooth start that ends in a breakdown. C and D cover the same two settings.

The fine print is in the force. In A and B, the text says: “Take f(x, t) to be identically zero.” That means no outside force, so the fluid is left alone. But C and D ask for a smooth start “and a smooth f(x, t).” There, an outside force is allowed, as long as it is smooth and well-behaved.

So the two halves of the prize do not mirror each other. One half takes the spoon away, and the other half hands it back.

Did ChatGPT Solve Navier-Stokes?

This was not the ChatGPT on your phone. OpenAI says it used an unreleased internal model. It ran a crowd of AI programs, called agents, that could talk to each other and run code.

The numbers in OpenAI's post are large. The group that cracked the problem had “on the order of 10,000 concurrent agents.” They reached their proof on Saturday, 5 September, “about 88 hours after the first agents were launched.” Along the way they sent each other 2.7 million messages, which is some group chat.

The proof starts with a fluid at rest and applies a smooth outside force. The fluid's energy stays finite the whole way, and even so, in a finite time, the flow blows up. OpenAI describes a swirl that spirals inward and gets stretched long and thin, “like spaghetti.” It says this settles statement C, and D as well.

A second AI model then spent 17 hours rewriting the proof in Lean. Lean is a programming language that checks a proof line by line. One job is left for people, Quanta notes: someone has to confirm that the statement Lean checked is the one Fefferman wrote. And peer review of the proof has not happened yet.

Two more details are worth a smile. OpenAI pointed its agents at every open Millennium Prize problem. It has reported a result for only this one, so the Riemann hypothesis lives on. And Quanta reports that OpenAI's Sébastien Bubeck put the computing bill at “several million dollars.” OpenAI also wrote: “We do not intend to claim the Millennium Prize for this result.” So it spent several million dollars to not collect one million. Nobody said the machines were good with money.

Why Do Some Mathematicians Call The Outside Force A Loophole?

Scientific American reports that most experts write the problem without that little f. They assumed the force did not matter, and that any blow-up would look the same with or without it.

Córdoba and his former student Luis Martínez-Zoroa saw an opening: could you break the equations using the force alone? Their method stacks up endless layers of well-behaved flows. Martínez-Zoroa calls it an “infinite cascade.” By 2023, Quanta reports, they had blown up Euler's friction-free equations this way. But their force came out messy, and the prize demands a smooth one, which is the gap the AI work closed. Córdoba, for the record, likes to joke: “I don’t use AI: I have Luis.”

In the proof, the fluid starts at rest, so every bit of its motion comes from the outside force. In coffee terms, the prize text allows a spoon, and this is a brilliant spoon. It stirs in a pattern that stays smooth even as the coffee under it breaks. OpenAI stresses that the blow-up comes from the fluid's own motion, and not from “putting in an infinite force by hand.”

That is fair, but it is not the question most experts had in mind. They want to know about coffee with no spoon: can a fluid that is left alone break on its own? Scientific American sums it up: “The Clay problem, as written, is solved.” The problem as many experts picture it is not. Some, the magazine warns, may say the win has “only been achieved through a loophole.” It adds that it is not clear the proof can be stretched to cover the no-force case.

There are hints, though. OpenAI says its agents also blew up Euler's friction-free equations with no outside force at all. That claim awaits review too. Tao weighed in on his blog, the day before OpenAI's news. He wrote that “it should even be possible to do without the forcing term.” That is a forecast about what should be possible, not a proof.

Who Solved It First, OpenAI Or The Mathematicians?

This part of the story is disputed, so here are only the reported facts.

Tristan Buckmaster is a mathematician at New York University. For about a year he worked with Levent Alpöge, who works at the AI company Anthropic. They pushed the Córdoba and Martínez-Zoroa method further with AI help. Scientific American gives the key date as 15 August. That day, they proved that Euler's equations blow up under a force. Buckmaster posted their results at 11:58 p.m. EDT on 7 September, and OpenAI's news followed about 12 hours later.

In his statement, Buckmaster alleged that word of their work had reached OpenAI, and suggested OpenAI may have gained from it. OpenAI denies this, and its post says it saw none of the pair's work until it was public. After an internal review, it adds, Buckmaster's prompts “could not have influenced the system in any way.” OpenAI does grant the pair priority for the forced Euler result.

Science has seen fights over credit before. Here, though, all sides agree that the key ideas belong to Córdoba and Martínez-Zoroa. Fefferman told Quanta they are the heroes of the story, and Buckmaster wrote that Martínez-Zoroa “deserves a Fields Medal.”

So, Is The Navier-Stokes Problem Solved Or Not?

It depends on which problem you mean, and there are three.

The problem as written now has a Lean-checked proof, covering statements C and D. The Clay Institute says the problem has “apparently been settled.” That word “apparently” is doing a lot of work. Clay adds that its review is “deliberately unhurried.” As of this writing, no prize has been awarded, and OpenAI says it will not claim one.

The problem as experts meant it asks whether a fluid with friction, left alone, can break down. Nobody has shown that it can, and nobody has shown that it cannot. Statements A and B remain unproved.

The problem in your cup has not changed at all. Planes fly and forecasts run on the same equations as last month. Even so, the result matters, because friction was meant to be the peacemaker, and the math now says it can lose. Quanta's verdict is a good one: “Turbulence is even weirder than it appears to be.”

So stir your coffee and watch the swirl die, as it always will. The mystery was never the mug, but whether a law as plain as F = ma could ever run out of answers. For a fluid left alone, nobody knows yet. Hand it the right spoon, though, and a proof that still awaits peer review says it can.