A Riemann hypothesis proof developed with substantial help from the AI model Claude is correct, according to the Lean and Mathlib communities. Levent Alpöge, a mathematician who works at Anthropic, released a formal proof library of about 17 million lines of Lean code on Friday, November 20, 2026. On Saturday, the communities behind the Lean proof assistant and the Mathlib mathematics library, endorsed by the Lean FRO and the Mathlib Initiative, said they had “full confidence” that the proof is correct. If that holds, one of the most famous open problems in mathematics has been solved, and by an unusual route: first as machine-checked code, with a readable paper still to come.
What the Riemann hypothesis is about
Bernhard Riemann stated the conjecture in 1859. It concerns the zeros of the so-called zeta function and is closely tied to how regularly the prime numbers are spread along the number line. Put simply, if the hypothesis is true, the number of primes below a given bound deviates as little as possible from the expected estimate. Over the decades, countless theorems in number theory have been proved on the assumption that the Riemann hypothesis holds. Since 2000, the Clay Mathematics Institute has listed it as one of seven Millennium Prize Problems, each carrying a one-million-dollar prize.
Alpöge claims even more. In his four-page note, he writes that the method also settles the generalized Riemann hypothesis, specifically for Dirichlet L-functions. The formal library contains five headline theorems, including the classical hypothesis and the generalization.
How the Riemann hypothesis proof came together with Claude
Alpöge describes the collaboration in unusually emotional language. He says he had done a lot of mathematics with the model in recent months, “but this time Claude far outpaced me near the finish line.” He set the direction himself: combine two quite different ideas from a simplified toy setting and carry them over to the right situation. Many of the technical steps, especially around automorphic forms on higher-rank groups, then came from Claude, an area in which Alpöge says he has little expertise of his own.
According to the note, the work ran roughly from early October to mid-November 2026. Alpöge says the timeline he prints was compiled by Claude. It starts on October 6 with a determinant modelled on the work of the mathematician Sergei Stepanov, whose argument settles the Riemann hypothesis over function fields, and ends on November 16 with the entry “GRH.” The central idea is a chain of so-called theta lifts that mimics, over the integers, the kind of multiplicity that Stepanov’s method can use over function fields.
What stands out is how openly Alpöge discusses the limits of his own understanding. He says he has a complete argument before having mastered the mathematics around it. He considered it inappropriate to publish a human-readable version: Claude produced a “book” of more than 250,000 pages that nobody could meaningfully read. So he released only the Lean code. He expects the proof to be “massively simplified” and plans to deliver a proper write-up.
What Lean checks, and what it does not
Lean is a proof assistant: a programming language in which mathematical statements and proofs are written down so precisely that a computer can check every step. Mathlib is the large, community-maintained mathematics library for it. A formal proof counts as correct if it compiles without errors, uses only the usual foundational assumptions (axioms), and actually proves the intended statement.
Those are exactly the points the community checked, according to its joint statement of November 21:
- The statement is right: the formulation of the Riemann hypothesis has been part of Mathlib for some time and has been vetted by experts in Lean and number theory.
- The code compiles: dozens of people built the library without errors and obtained the same raw proof export.
- It proves the right statement: a separate file that imports only Mathlib isolates the statement. Two independent tools confirmed that Alpöge’s theorem is exactly that statement.
- The proof is valid: several proof checkers, including Lean’s official kernel and the con-leche and con-ron checkers, accepted the exported proof. The latter two come with a formal proof that they cannot accept a false proof.
One safety detail matters: the checks ran on an inert export of the proof. Any malicious code hidden in the Lean sources could not have influenced the checkers. With 17 million lines of machine-generated code, that is not a theoretical concern.
The numbers behind the verification
The scale of the effort is visible on the verification page of Andrew Sutherland at MIT, who says he ran his checks with Claude Code. All 175,187 modules of the library compile without errors. The five headline theorems rely only on Lean’s three standard axioms. The exported proof is just over 61.5 gigabytes. An export built independently in a sandbox without network access was byte-for-byte identical to the published one.
The independent kernel nanoda, written in Rust, needed just under 28 minutes on 80 threads for the full export. con-leche accepted 1,468,418 declarations. According to the page, two runs of Lean’s default kernel on the full export were still in progress as of 05:35 CET on Sunday, while the module-by-module replay through Lean’s kernel had already finished without a single rejection. Michael Rubinstein of the University of Waterloo also reported that his four checks finished with no errors. He is also working on a 164-page guide to the proof for the zeta function, written with the help of Claude, Codex and Gemini. That guide, however, still takes the core step, an inequality between weighted sums over primes, from the Lean code.
Timing and context: the IPO and the AI race
The release came one day after Anthropic’s stock market debut. Alpöge addresses this in his note: the timing is “horrible,” and he would have preferred to finish earlier. He says he asked the company’s leadership whether he could announce at all and that there was a constraint to respect. He explicitly calls the suspicion that the announcement was timed to the IPO false. We covered how the $83 IPO price came about at the time of the listing.
The case is part of a run of AI-assisted results in mathematics. OpenAI reported several results this autumn, which we covered in our piece on the OpenAI math proof on ETH and Kakeya. Alpöge explicitly praises that work in his note. Google DeepMind also uses AI agents on hard problems, for example in its analysis of RSA-1024. The difference here: for the Riemann hypothesis there is not just a claim, but a complete, machine-checked proof.
What remains open
Formally, the proof is secure in the view of the Lean community. Mathematically, the work is only beginning. Nobody, Alpöge included, can currently explain the full proof in human language. Experts will want to establish which ideas are genuinely new, whether the argument can be shortened, and what it means for other conjectures, such as the extended Riemann hypothesis that Alpöge names as his next target. The Lean community says it looks forward to the work needed to explain the new ideas.
It is also unclear how the Clay Mathematics Institute will treat a proof that initially exists only as code and in which an AI played a major role. Its rules require, among other things, publication and a waiting period with broad acceptance by the mathematical community. There was no statement from the institute as of Sunday morning.
Our take
For mathematics this is a turning point, and for the AI debate a test case. It shows that formal verification can solve the trust problem around AI results: whether a human or a machine wrote a proof does not matter for its validity if independent checkers accept it. But it also exposes a new gap. A result can be correct without anyone understanding it. Whether it becomes knowledge will be decided over the coming months, as people retrace the path that Claude and Alpöge took.




