AI Research2026-11-22Updated 2026-11-22

Riemann Hypothesis Proof Certified by Lean: Claude's Role (Nov. 2026)

TL;DRLean FRO and the Mathlib Initiative certified Levent Alpöge's formal proof of the Riemann Hypothesis on November 21, 2026. The 17-million-line Lean library was released the previous day, after work that Alpöge describes as a collaboration with Anthropic's Claude. Independent checks compiled 175,187 modules and matched the proved result to Mathlib's vetted statement. This establishes the validity of the formal statement and proof. It does not replace the mathematical community's work to understand the argument, and it is not a Clay Millennium Prize award.

By Cameron Voss and West Mao

17 millionlines of Lean code
175,187modules compiled without errors
5headline theorems checked

The joint statement came one day after Alpöge released his proof. Its central sentence leaves little room for ambiguity:

"We hereby certify that both the statement and the proof of the Riemann Hypothesis are correct."

Lean FRO and Mathlib Initiative, November 21, 2026

The Riemann Hypothesis concerns the zeros of the Riemann zeta function and their connection to the distribution of prime numbers. Bernhard Riemann posed it in 1859. Alpöge's result arrives as executable mathematics rather than a conventional paper: a vast Lean library whose claims can be checked by a computer.

What did Lean and Mathlib actually certify?

They certified that the submitted library proves the intended mathematical statement and that its proof is valid under Lean's formal system. That distinction matters. The organizations did not certify a prose manuscript or announce that every question surrounding the result had been settled.

The statement names Alpöge and congratulates him. It does not name Claude as an author or claim that Anthropic produced the proof independently. Claude's involvement comes from Alpöge's own account and from the tools used during verification.

How was the proof checked?

The community addressed four ways a formal proof can fail: the statement could be wrong, the code might not compile, the proved theorem could differ from the advertised one, or the proof itself could be invalid. All four checks passed.

QuestionWhat the verification established
Is the statement correct?The result uses Mathlib's RiemannHypothesis definition, which Lean and number theory experts had vetted.
Does the library compile?Dozens of people reported error-free compilation and reproduced the same raw proof export.
Is it the same theorem?comparator and con-leche --compare matched Alpöge's result to the Mathlib statement.
Is the proof valid?Lean's official kernel, con-leche, con-ron and independent checkers accepted the exported proof.

MIT mathematician Andrew Sutherland published a more detailed record. His checks, run with Claude Code, compiled all 175,187 modules with zero errors. The five headline theorems depend only on propext, Classical.choice and Quot.sound, Lean's three standard axioms. The independent nanoda and con-leche kernels accepted the complete exported proof, while lean4lean accepted every compiled module.

The checkers were not asked to trust arbitrary executable code in the source repository. They examined an inert proof export. The Lean statement also notes that con-leche and con-ron carry formal proofs, relative to set theory, that they cannot accept a false proof.

Did the proof also cover the Generalized Riemann Hypothesis?

Yes, for Dirichlet L-functions. Sutherland's five checked theorems include the Riemann Hypothesis and results placing the relevant zeros of Dirichlet L-functions on the critical line. The scope covers all moduli and characters, including the usual formulation for primitive characters.

That should not be shortened to "every form of GRH." The verification did not cover Dedekind zeta functions, Hecke L-functions or automorphic L-functions in general. The narrower description is the accurate one.

What did Claude contribute?

Alpöge's release note describes the proof as work done with Claude and says the model pulled far ahead of him near the finish. The available certification does not break down which ideas came from the mathematician and which came from the model. Nor does it identify a particular Claude model version.

Claude also appeared on the verification side. Sutherland states that he used Claude Code to run his checks. Those checks still relied on Lean and independently implemented kernels to decide whether the proof was valid. Claude Code operated the process; it did not serve as the authority that accepted the mathematics.

DeepSeek for technical work

Working with AI on code, formal methods or technical analysis? Explore DeepSeek model Coding Plans to check the available purchase modes and current stock.

Why does this result matter for AI-assisted research?

The strongest evidence here is not a model claiming to have solved a famous problem. It is an artifact that other people can compile, export and check with separate tools. That makes the claim far easier to test than a plausible-looking proof in a chat window.

The result also exposes a different bottleneck. Producing formal arguments at this scale is one task; understanding them is another. Alpöge credits Claude with substantial assistance. Even so, he is the named author, and separate proof checkers supplied the verification.

Does formal certification mean mathematical review is finished?

No. Machine verification and human understanding answer different questions. A kernel can establish that a conclusion follows from the supplied definitions and axioms without giving readers a compact explanation of the new ideas.

The Lean and Mathlib statement explicitly looks ahead to work that will explain those ideas. For a 17-million-line library, that is a substantial job. The proof has passed an unusually demanding correctness check, but mathematicians still need to digest the argument and place it within number theory. Formal certification is not the same thing as a journal review or a Millennium Prize decision.

Sources

  1. Lean FRO and Mathlib Initiative: "Lean verification of the Riemann Hypothesis" (November 21, 2026)
  2. Andrew Sutherland, MIT: "Verification files for the Lean proof of the Riemann Hypothesis" (verification record)