Skip to content
sobolewski.proAI today, every eveningSubscribe

Once a day. Unsubscribe with one click. Privacy policy

Navier-Stokes proof: finite-time blow-up claimed, written with Claude

A Navier-Stokes proof of finite-time blow-up: Alpöge and Buckmaster posted a paper with a Lean formalization, written with Claude. Here is what we know so far.

4 min read

This article in other languages:Polski

In this article
  1. What does the Navier-Stokes proof claim?
  2. What does the proof look like?
  3. What was Claude’s part in it?
  4. What should you watch for?
  5. FAQ

On November 20, 2026, Levent Alpöge and Tristan Buckmaster posted a 79-page paper with a Navier-Stokes proof: they claim that the three-dimensional Navier-Stokes equations have solutions that blow up in finite time. Whether that can happen is one of the seven Millennium Prize Problems, and the authors write that they used Claude models throughout the work. There is no independent confirmation yet.

What does the Navier-Stokes proof claim?

The equations describe the flow of fluids such as water and air, which is how the Clay Mathematics Institute presents them. The open question was whether a flow that starts out smooth stays smooth forever, or whether a singularity (a “blow-up”) can form in it after a finite time, meaning a point where the quantities describing the flow grow without bound.

Alpöge and Buckmaster answer that a singularity can form. The abstract says there exist smooth initial data, both in the whole space and on the torus, for which the solution with no external force at all stops existing as a smooth solution in finite time. According to the authors this disproves statements (A) and (B) in the official formulation of the Millennium Problem and so proves statements (C) and (D).

The no-force caveat matters. The authors’ earlier papers, which they cite, dealt with equations that had a force added, and the Millennium Problem asks about the case without one.

What does the proof look like?

It has two parts. The first shows that if a certain special solution exists, one that repeats itself at ever smaller scales (a “discretely self-similar” solution), then you can get from smooth initial data to a singularity. The second shows that such a solution exists near an approximate one found numerically, provided the approximate one satisfies a list of specific inequalities. The authors call that package a “certificate”.

The approximate solution was found with Newton’s method. It is not axially symmetric: it stays unchanged under a rotation by 1/19 of a full turn (more of a curiosity, but it shows nicely that nobody would have guessed this shape on paper ;)).

Terence Tao described the result on his blog the same day. In his view the strategy, building an approximate solution numerically and getting from it to an exact one, is one of the expected paths to this result.

What was Claude’s part in it?

The paper has a separate “AI statement”. In it the authors write that they used Anthropic’s models the whole time and nothing else: “We used Anthropic models (Claude) throughout this work, and no others”. Which version, they don’t say.

From the description, the agents did not start from scratch. The authors gave them “a curated dossier of potential leads”, and guidance after that. The paper also builds on earlier, partly unpublished work by both mathematicians on computer-assisted proofs.

The order was the reverse of what is usual in mathematics. The Lean proof came first, and only then was a first version of the text generated from it, which the authors rewrote with Claude under their own direction. According to the repository README, the Lean library is about 2,270 files and 600,000 lines. In the acknowledgments the authors thank Anthropic for supporting the collaboration the paper grew out of.

It is the second announcement of this kind in one day with Alpöge’s name on it. A few hours earlier he announced a proof of the Riemann Hypothesis found with Claude. Tao puts three results side by side in a single post, from the most to the least “digested”: a singularity for the Euler equations (55 pages, by Alpöge, Buckmaster and Matei Coiculescu, with Claude and Codex used there), a singularity for the Navier-Stokes equations, and last the Generalized Riemann Hypothesis.

What should you watch for?

The paper and the repository show three things that are still missing:

  1. The certificate. It has not been published. The paper gives only its hash and says it will be released soon. Without it the second part of the proof cannot be checked.
  2. The computations. That the certificate satisfies the required inequalities is checked by a separate program based on interval arithmetic (128 bits of precision, the Arb library). That program is not part of the Lean formalization, and it is not in the repository either.
  3. The formalization itself. According to the README, a patch was applied to it on November 20, and the patched package was not rebuilt before publication.

Tao writes that the paper is “technically human-readable” and that more work is needed to produce a properly written proof. Nothing in his post suggests he has checked it himself.

So there is more to check here than with the Riemann Hypothesis, where there is no text to read at all, but it is still work for people. A Lean formalization as the main argument is something we also saw in October, when an OpenAI model was said to have disproved the exponential time hypothesis.

FAQ

Has the Navier-Stokes Millennium Problem been solved?

We don’t know yet. The authors say it has, but as of November 20, 2026 there was no independent confirmation, and the Clay Mathematics Institute still lists only one of the seven problems as solved, the Poincaré conjecture. Each carries a $1 million prize.

Did Claude prove it on its own?

The paper does not say that. The authors describe a collaboration: Claude agents were given prepared leads and guidance, and the text was rewritten many times under the mathematicians’ direction.

Where can I see the proof?

The paper and the Lean code are in Tristan Buckmaster’s GitHub repository. The certificate and the program that checks it are not there.

  • AI
  • Anthropic
  • Claude
  • science

AI today, every evening

The day's biggest AI news and new posts from the blog, every evening. Short, with sources.

Once a day. Unsubscribe with one click. Privacy policy