Friday, 20 November 2026

Riemann hypothesis solved by AI? A 14-file proof faces its test

An Anthropic mathematician says he and Claude cracked math's most famous open problem. Nobody has confirmed it yet, and the proof exists only as computer code.

A chalkboard with faint curves and a vertical line behind a laptop on a desk

The short version

  • Anthropic mathematician Levent Alpöge says he found a proof of the Riemann hypothesis while working with the Claude AI model.
  • The proof was released only as Lean computer code in 14 archive parts, with no written paper so far.
  • No independent check of the full proof has finished, so the 1859 problem is not officially solved.
  • The Clay Mathematics Institute offers a $1 million prize for a correct proof of the hypothesis.
Quick read · 1 min

A mathematician at Anthropic says he and the Claude AI model have proved the Riemann hypothesis, the best-known unsolved problem in math. Levent Alpöge's website now hosts a four-page note and 14 archive files of computer code that are said to contain the full argument.

It is a claim, not a confirmed result. There is no written paper. The proof exists only in Lean, a language that lets a computer check every logical step. An independent project has confirmed the files match the release, but its test of the proof itself was still running.

  • The problem dates from 1859 and concerns how prime numbers are spread out.
  • A correct proof carries a $1 million prize from the Clay Mathematics Institute.
  • Nothing changes for banking or internet security either way.

Experts expect code checks to settle the matter quickly. A readable write-up is promised later.

Is the Riemann hypothesis solved by AI? Not yet, and maybe not at all. But on Friday a mathematician who works at Anthropic posted a short note saying that he and the company’s Claude model have proved the most famous open problem in mathematics, a question that has beaten every human who tried since 1859.

If it holds up, an AI model will have helped do something the best number theorists alive said they were nowhere near doing. If it doesn’t, it will be one more loud AI math claim that fell apart under checking. Right now nobody outside the author’s circle can say which, and the way the proof was released explains why.

01

Is the Riemann hypothesis solved by AI, or only claimed?

It is only claimed. The claim comes from Levent Alpöge, a mathematician employed by Anthropic, the company that makes Claude. A four-page note titled “Remark” appeared on his personal website, written in the first person. Its author says he believes that, working with Claude, he has found a proof.

There is no published paper and no peer review. One report points out that the note carries no author name and that Alpöge had not shared it on social media, although other coverage treats it plainly as his statement. Either way, a note on a website is where a proof starts, not where it ends.

The note goes further than the classic problem. It says the same method also proves the Generalized Riemann Hypothesis, a much wider version that covers a whole family of related functions known as Dirichlet L-functions. The author adds that an even broader form, the Extended Riemann Hypothesis, is something he hopes to reach with more time. That part is future work, not a claim.

Server racks with blue status lights in a data center
02

What the Riemann hypothesis says, in plain English

The Riemann hypothesis is a guess about prime numbers, the numbers that divide only by themselves and one. There are 25 primes up to 100 and 168 up to 1,000. They look scattered, with gaps that grow and shrink with no clear rhythm. Bernhard Riemann suggested in 1859 that there is a deep order underneath.

He tied the primes to something called the zeta function. Certain inputs make that function equal zero, and those inputs are called zeros. The hypothesis says every one of the interesting zeros sits on a single vertical line, where the real part of the input equals one half. If that is true, the primes are spread out about as evenly as they possibly could be, and estimates of how many primes exist below any number can only miss by a small, known amount.

Computers have already checked the first ten trillion zeros, and all of them sit on the line. That is not a proof. One zero off the line, anywhere past the checked range, would sink the whole idea, which is why mathematicians want an argument that covers every case at once.

The stakes inside math are large. Thousands of results in number theory have been proved only on the assumption that the hypothesis is true. It is one of the seven Millennium Prize Problems, and the Clay Mathematics Institute has a $1 million reward waiting for a correct proof.

03

Why this proof came as 14 code files and not a paper

The proof was released only as Lean code because, the note says, the argument is too long and too unfinished to write up for human readers yet. Lean is a programming language built for math. A proof written in it can be checked by a computer one logical step at a time, so a mistake can’t hide behind a confident sentence.

Alpöge’s site hosts the code as 14 separate compressed chunks that have to be joined back together. A readable version is promised for later. A footnote in the note jokes that the informal, AI-written draft material ran to more than 250,000 pages.

The note describes Claude as far more than a typing aid. It calls the work a joint effort, says the model raced ahead of the author near the end, and credits Claude with the central trick. In rough terms, the approach borrows a method that already works in a simpler mathematical setting and replaces a missing ingredient with a long chain of steps called theta lifts. A timeline in the note, which it says Claude put together, has 20 steps and runs from early October to mid-November, with the last step dated November 16.

The author is frank about the rough edges. He warns that the argument may include odd choices, that nobody has looked for a shorter route, and that some descriptions of Claude’s part are paraphrased from an AI-written summary and may not make sense.

An open notebook and pencil on a table in a library reading room
04

What has been checked so far, and what hasn’t

Very little has been confirmed so far. An independent project reported within hours that its copy of the files matched the release, and that the statements it wanted to test could be written using Lean’s standard math library. The same project said its check of the proof itself was still running.

Two questions decide everything. Does the whole archive build without errors and without leaning on extra assumptions? And does the final theorem say exactly what the Riemann hypothesis says? The note claims the formal proof is complete and uses no added assumptions, but it doesn’t give build instructions or list the basic axioms it relies on. One early look at the top-level files found that the classic theorem targets the same definition of the hypothesis that Lean’s math library already contains, which is a good sign and nothing more.

A verdict should not take long. Problems can crop up in Lean, but experts can check fairly quickly whether the hypothesis was coded properly and whether the proof is correct.

05

What changes for you if the proof holds

Your bank login and your internet speed stay exactly the same. A correct proof has no immediate real-world use, and it would not upend global banking. Confirming the hypothesis tells mathematicians that a long-trusted assumption was right all along.

The bigger meaning is about AI. Alpöge has been in the news since July, when he posted a Claude-made solution to another well-known problem, the Jacobian conjecture. On the same day as the Riemann note, he and New York University mathematician Tristan Buckmaster posted a separate claimed result on the Navier-Stokes equations, which describe how fluids flow. That one is also a Millennium Prize Problem, and it is also unverified.

If either survives, it would show AI models doing original work at the very top of a field, not only passing exams. Alpöge’s note admits that some mathematicians will feel a sense of loss. It also calls the timing terrible and accepts that there is a cynical way to read it. He insists that reading is wrong and says going public now was his own call, agreed with Anthropic’s leadership and a few trusted mathematicians.

06

Quick answers

Has the Riemann hypothesis been proven?

Not officially. A proof made with Claude has been claimed and posted as Lean code, but independent checking of the full argument had not finished at the time of writing.

Who gets the $1 million prize if AI solved it?

Nobody has been awarded anything. The Clay Mathematics Institute offers $1 million for a correct proof, and this claim has not been confirmed.

What is Lean?

Lean is a programming language for mathematics. A computer can use it to check each logical step of a proof, which makes errors much harder to miss.

You're reading the quick version.