The Millennium Problems Are Falling. The Argument Is About What Fell.

Two months, four dates, and the boundary of what a machine can contribute to mathematics moved further than it had in the previous thirty years.

8 September — the fluid result
OpenAI publishes a proof of finite-time blowup for the Navier–Stokes equations in three dimensions, with a Lean-checked formal proof. The run used a fleet of autonomous agents on an internal model. The claim is a negative resolution: the equations can, for some initial data, stop being smooth.
18 September — the second proof
OpenAI says the same programme produced a proof for quantum chromodynamics, the theory of the strong force.
16 October — formalisation finished
The Yang–Mills existence argument is fully formalised in Lean, the proof assistant that checks a derivation against explicit logical rules rather than against a reader’s intuition. Anthropic publishes its own refutation of the Exponential Time Hypothesis the same day.
26 October — the announcement
OpenAI announces three results together: local smoothing and the Kakeya conjecture in every dimension, a refutation of the Exponential Time Hypothesis, and an existence proof for pure Yang–Mills theory. Two arrive with public Lean proofs. The physics one does not.

What was actually claimed

The precision matters, because the headlines blurred it.

Navier–Stokes is one of the seven Millennium Prize problems posed by the Clay Mathematics Institute in 2000, each carrying a million dollars. The question is whether solutions to the three-dimensional equations always stay smooth, or whether they can blow up in finite time. OpenAI's paper is titled Finite time blowup for Navier–Stokes: it claims the second outcome, for some initial data.

Yang–Mills is another of the seven. That problem has two halves. The first is to show the theory exists as a rigorous mathematical object. The second is to show it has a mass gap — that the lightest particle the theory predicts has a mass strictly above zero, which is why the strong force is short-range.

Millennium Prize problems, each carrying one million dollars, posed in 2000

of them touched by AI results this autumn: Navier–Stokes and Yang–Mills

of the two Yang–Mills halves claimed — existence. The mass gap is untouched

million dollars awarded. No result has been accepted for a prize

OpenAI's October claim is the existence half. It is not the mass gap. The Clay problem requires both, so no prize is claimable on what has been announced, and OpenAI's own statement describes the result as an existence proof rather than a solution to the Millennium problem.

Existence1 halves claimed

OpenAI claims this on 26 October — the theory exists as a rigorous object

Mass gap0 halves claimed

the lightest particle has mass above zero; untouched by any AI result so far

The Clay problem needs both halves. One has been claimed, none has been awarded.

The dispute around the fluid result

The Navier–Stokes claim is the more contested of the two, and not because anyone has found an error.

Critics — Scientific American among them — argue that OpenAI proved blowup for a version of the equations that is not the one the Clay problem asks about, and that the difference is a loophole rather than a detail. Three mathematicians then published a proof that OpenAI's method cannot be extended to the full problem, which closes the loophole in the other direction: the technique works, and it cannot reach the question as posed.

There is a priority dispute in the background as well. Mathematicians Levent Alpöge and Tristan Buckmaster had been working publicly on related Euler and Navier–Stokes results; OpenAI has said its effort was prompted by rumours of that work, and that the two proofs differ substantially. The New York Times reported the clash in September.

None of that makes the result worthless. A machine-produced, machine-checked proof of blowup for a well-defined variant is a real mathematical object. It is simply not the same thing as a solved Millennium problem, and the gap between those two sentences is where most of the argument lives.

Anthropic's answer, and why it matters

On 16 October, Anthropic published a refutation of the Exponential Time Hypothesis — the conjecture that certain computational problems genuinely require exponential time — found by an internal Claude model. The paper credits the model, and ships a Lean 4 proof checked by a second, independent kernel.

That last detail is the important one. A second kernel means the proof was not merely produced by a machine and believed by its producer; it was checked by different software implementing the same logical rules. Independent checking is what turns a claim into a result, and it is the part of this story that will outlast the announcements.

The bottleneck is no longer generation

For most of computing history, the hard part of a mathematical result was finding it. That is no longer obviously true.

Produce
A model or a fleet of agents generates a candidate proof. This is now cheap enough to attempt at scale, repeatedly, on variants.
Formalise
The argument is translated into Lean. This is where hand-waving dies: every step must typecheck against explicit rules.
Release
The proof and the formalisation are published. OpenAI released two of the three October results and withheld the physics one, pending expert review.
Review and accept
Mathematicians read it, try to break it, and either accept it or find the gap. This is now the slow step, and no amount of compute shortens it.

Peter Woit, the Columbia mathematician whose blog is a reliable sceptic, described the Yang–Mills situation bluntly: a result held privately while experts are assembled to check it is not yet mathematics in the sense the field means. That is a fair standard, and it applies to the good results too.

Why foundations are worth the fuss

Mathematics and physics are not two subjects among many. They are the layer the others are built on.

When the mathematical footing of a theory of matter becomes tractable, the consequences arrive later and elsewhere: in materials designed rather than discovered, in drug targets found by simulation before a lab is booked, in chemistry that stops being a matter of trial and error. Anthropic's public bet on biology and medicine is not a pivot away from its models; it is a bet that better reasoning about physical systems converts into biological ones.

results OpenAI announced together in October: Kakeya, the ETH refutation, and Yang–Mills existence

shipped with public Lean proofs; the third is held pending review

independent kernel check on Anthropic’s ETH refutation — the model of what verification should look like

Millennium problems7 open

posed in 2000; none has been accepted as solved by this autumn’s results

Touched by AI results2 open

Navier–Stokes (disputed framing) and Yang–Mills (existence half, unreleased)

Two problems touched is not two problems solved. The distance between those sentences is the whole story.

The author's take

This is opinion, not reporting. It is the argument made in the video accompanying this article.

AI is getting so smart that it’s solving math’s most difficult problems, one by one. I wouldn’t be surprised in a year or two if all the remaining million-dollar math problems are solved by AI. I’m reminded back in the day when AI basically solved chess by beating the best human. Now AI is beating the best mathematicians, and it’s not even close. But I think this time is different, because AI is now a generalized intelligence, meaning it covers almost all areas and disciplines. So the potential and possibilities are nearly endless.
When you think about math and physics, it’s the building blocks, the foundation for a lot of how we understand the world. So as AI helps solve the hardest math and physics problems, it’s going to lead to countless discoveries in fields that are built on math or physics. This is one of the reasons why Anthropic is so bullish on biology and medicine, and why they’re building their own lab efforts. I think we’re going to see countless new discoveries in material sciences too — we’ll figure out new ways to make things, build things.
Basically, AI is helping us figure out how the world and our body work, and that understanding is what will lead to mind-blowing discoveries at a rate never seen in history. This is why this topic is so interesting. We’re at a big turning point. Things are speeding up. I’m not saying there aren’t going to be challenges — there are. But the pace of new discoveries is speeding up, and it’s just the beginning.

The sober version of that argument is narrower and harder to dismiss. Machines have moved from assisting mathematicians to producing results that a proof assistant can check, on problems that were open for decades. The verification is slow, some claims will not survive it, and a prize nobody has been awarded is not a prize won. But the direction is no longer in doubt.

Source: OpenAI, "On the Navier–Stokes Millennium Prize Problem" (8 September 2026) and "Working with academic communities to build scientific understanding" (26 October 2026); Anthropic, exponential time hypothesis refutation (16 October 2026); Scientific American, "Did OpenAI solve the wrong Navier–Stokes problem?" (21 September 2026); Quanta Magazine (8 September 2026); The New York Times (10 September 2026); Clay Mathematics Institute, Millennium Prize problem statements.

Contact: [email protected]