The Lean and Mathlib communities say Levent Alpöge's proof of the Riemann hypothesis, found with Claude, is correct. In a joint statement they write: "We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct." Alpöge works at Anthropic, which makes Claude.

The Riemann hypothesis, first posed in 1859, is a famous open question about how prime numbers are spread along the number line. Alpöge posted his proof on Friday, Nov. 20, as a library of Lean code, a language that lets a computer check each step.

What was certified

The statement is dated Nov. 21 and is endorsed by the Lean FRO and the Mathlib Initiative, which support the Lean language. It reports four checks:

  • The statement of the Riemann hypothesis is correct. It is part of Mathlib, Lean's main math library, and the statement says it has been vetted by experts in Lean and number theory.
  • Alpöge's library compiles. The library has 17 million lines. The statement says dozens of people reported compiling it without errors and getting the same raw proof export file.
  • The code proves that statement. Two tools, comparator and con-leche, confirmed that the theorem proved in Alpöge's library is exactly Mathlib's statement of the Riemann hypothesis. The codebase depends directly only on Mathlib.
  • The Lean proof is valid. The exported proof was checked by several proof checkers, including Lean's official kernel and the con-leche and con-ron checkers.

MIT mathematician Andrew Sutherland, who ran his own checks, told Scientific American: "Honestly the chance that anything is fishy here is approximately zero."

What this does and does not show

This is a formal verification of the Lean proof. The statement does not describe journal peer review, and it says the community looks forward to work that "will help explain the new ideas in this proof," so how the argument works is still to be explained.

Sutherland's verification page adds limits. When last updated early on Nov. 22 (04:35 UTC), it listed his own Lean-kernel replay and his own lake comparator run as still in progress. Those are his separate runs, not the checks the Lean statement reports. It says a passing check relies on the checking kernels and on Mathlib's definitions being right. Sutherland ran his checks with Claude Code. It also says the "generalized" part it checked covers Dirichlet L-functions only, not Dedekind, Hecke or automorphic L-functions.

We found no mathematician disputing the result. The Guardian reports that some mathematicians are uneasy about wealthy private companies spending heavily on AI to solve math problems.

Who did what

Alpöge says Claude helped him over about six weeks and that "this time Claude far outpaced me near the finish line," the Guardian reports. The proof is credited to Alpöge, working with Claude. He posted it a day after Anthropic's Nasdaq debut and wrote that "the precise timing is horrible." For the company's numbers, see our IPO filing story.

What this means for you

The claim is no longer just a claim: two Lean organizations say the formal proof checks out, and an independent MIT mathematician agrees. What is left is for mathematicians to explain and review the ideas. We found no source on practical effects, so we say nothing about them. For another lab's math claims, see OpenAI's math and physics claims.

What we checked

We read the Lean and Mathlib statement, Sutherland's verification page, the Guardian and Scientific American reports, and Alpöge's own statement on Nov. 22. We did not run the Lean code ourselves.