Skip to content
sobolewski.proAI today, every eveningSubscribe

Once a day. Unsubscribe with one click. Privacy policy

Riemann Hypothesis proof: Anthropic mathematician says he found it with Claude

A Riemann Hypothesis proof, found with Claude and formalized in Lean: that is the claim from Levent Alpöge of Anthropic. Here is what we know and what we don’t.

4 min read

This article in other languages:Polski

In this article
  1. What do we know about the Riemann Hypothesis proof?
  2. What was Claude’s part in it?
  3. Has the proof been confirmed?
  4. Why now?
  5. FAQ

On November 20, 2026, Levent Alpöge, a mathematician who works at Anthropic, posted a four-page note on his website saying that he and Claude have found a Riemann Hypothesis proof. The Riemann Hypothesis has been open since 1859. The note comes with a Lean formalization, but nobody independent has confirmed it yet.

What do we know about the Riemann Hypothesis proof?

So far there is the note and the code. The note opens carefully: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” Then it goes further than that sentence promises. According to Alpöge, the same method also settles the Generalized Riemann Hypothesis (GRH), and “with more time” he expects it to cover the Extended one too.

The hypothesis itself says that all the “non-obvious” zeros of the zeta function have real part equal to 1/2. That is how the Clay Mathematics Institute describes it, and Clay put it on its list of Millennium Prize Problems. In practice it is about how precisely the distribution of prime numbers can be predicted.

A proof in the usual sense, a paper you can read, does not exist. Alpöge writes that the argument is “incredibly long”, and that the fully detailed document he generated for his own amusement ran past 250,000 pages. He decided that posting that kind of “slop” would be inappropriate, so he released only the Lean: an archive split into 14 parts, linked in a footnote of the note.

The author is not a random person. On his website Alpöge lists himself as an Anthropic employee and a junior fellow at the Harvard Society of Fellows, and his most recent listed paper resolves Hilbert’s tenth problem for all number fields (with Manjul Bhargava, Wei Ho and Ari Shnidman).

What was Claude’s part in it?

A large one, according to the note, and in the part of the argument the author has not yet mastered himself. Alpöge describes steering the work to the point where two different ideas from a simplified “toy setting” had to be combined, and then the model got ahead of him: “Claude far outpaced me near the finish line.”

He gives one example which, in his view, is enough to make mathematicians “fall out of their chairs”. The argument descends from Stepanov’s method for function fields, where multiplicity is the key ingredient, and multiplicity is hard to find over the integers. The way around it is supposed to be a chain of so-called theta lifts, and Claude suggested that for a von Mangoldt sum of length X the chain should have length √X.

The note also contains a 20-item timeline that Claude put together. The first entries are dated October 6, 2026, most of the final steps November 15, and the last one, titled just “GRH”, November 16. Alpöge warns that where Claude was the expert, he is paraphrasing a summary that “might make zero sense”, and he points readers to the Lean code. The note does not say which version of the model he used.

We saw a similar story on a smaller scale a few days ago, when Claude solved two open problems in black hole physics.

Has the proof been confirmed?

Not yet. Scientific American covered the claim the same day and quotes no independent mathematician. It says that issues can arise in Lean, but that it shouldn’t take experts long to verify whether the hypothesis is properly coded and whether the proof is correct.

Kit Yates, on his Substack, lists what exactly has to be checked:

  1. Whether the theorem proved in Lean really is the Riemann Hypothesis.
  2. Whether the definitions in the code faithfully match the standard mathematical objects.
  3. Whether no extra assumption slipped in along the way.

That is far less work than reading 250,000 pages, but it is still work for people. Alpöge admits as much: he is presenting a complete argument before having mastered the mathematics around it, and he expects the proof to be greatly simplified. He promises a proper write-up and talks.

A Lean formalization as the main evidence is something we already saw in October, when an OpenAI model was said to have disproved the exponential time hypothesis, also with Lean code and an independent group checking it.

Why now?

The author himself calls the timing “horrible”. He writes that everything came together late on a Sunday night, that he asked “Dario” whether he could announce the result, and that “there was a constraint to respect”. He does not say what it was.

For context: Anthropic priced its public offering on November 18 and closed it on November 20. Some people commenting on Hacker News connect those dates. Alpöge answers that in advance: the “cynical view” of the timing is false, he writes, and he could not bear to wait any longer.

There is no post about the proof on anthropic.com/news (as of the afternoon of November 20, Polish time), so the author’s note remains the only source.

FAQ

Has the Riemann Hypothesis been proved?

We don’t know yet. Levent Alpöge says it has and has released a Lean formalization, but as of November 20, 2026 there was no independent confirmation.

What is Lean and why does it matter here?

Lean is a programming language that checks every step of a proof. If the code compiles and the theorem at the end is a correctly stated Riemann Hypothesis, reviewers don’t have to read the whole argument line by line.

Where can I see the proof?

The footnote of the note links to the 14 parts of the archive with the Lean code. A version written for humans does not exist yet; the author says it will come later.

  • 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