Levent Alpöge, a mathematician who works at Anthropic, says he used Claude to prove the Riemann Hypothesis and its generalized form, and he has released that proof only as Lean code, with no human-readable paper. Whether the claim holds now depends on other mathematicians reconstructing an argument that Alpöge says he has not yet fully mastered himself.
Levent Alpöge Released His Riemann Hypothesis Proof as Lean Code Only
Alpöge published the proof on 20 November 2026 as a Lean archive split into 14 downloadable files, alongside a four-page statement on his website. In that statement he says that, working with Claude, he found a proof of the Riemann Hypothesis that extends to the Generalized Riemann Hypothesis, and that the two formalized the argument in Lean without assumptions. Lean is a programming language and proof assistant whose kernel checks each step of a formal proof mechanically. Alpöge wrote that generating a fully detailed written document from the formalization was routine, and a footnote says the one he produced ran past 250,000 pages. He judged that posting it would be inappropriate because no one could read it, so he presented only the Lean. The BBC reported that the work used a non-public version of Claude, and Terence Tao described it as generated with an internal AI model.
Alpöge Says Claude Outpaced Him on the Mathematics the Proof Depends On
Alpöge’s statement describes a division of labour in which he chose the strategy and Claude carried the technical execution. He says he steered the work toward combining two ideas he had seen succeed in a simpler toy setting, and that Claude then far outpaced him near the finish. He also writes that he personally has little mastery of automorphic forms on higher-rank groups, the area where much of the final argument sits. His own summary of the proof’s steps comes with a warning that the parts where Claude was the expert are paraphrased and might make no sense, and that the Lean code is the authority. A timeline Claude produced for him puts the first key step on 6 October 2026, and the BBC reported the full effort took about six weeks. Alpöge acknowledged the unusual position directly, writing that he is presenting “a complete argument before having mastered” the mathematics around it. He says he intends to understand the argument, simplify it and write it up properly, and he expects much shorter proofs to follow.
Terence Tao Ranked the Riemann Claim the Least Digested of Three AI-Assisted Results
Mathematician Terence Tao listed three results announced on 20 November 2026 in order of how well understood each one is. First came a proof of finite-time blowup for the three-dimensional incompressible Euler equations by Alpöge, Tristan Buckmaster and Matei Coiculescu, written in human-readable form and formalized in Lean. Second came a Navier-Stokes blowup result by Alpöge and Buckmaster that Tao called technically human-readable but still in need of a professionally written proof. Last came the Generalized Riemann Hypothesis proof, which Tao described as an enormous Lean-only proof. He pointed readers to a thread on the Lean Zulip chat as the place where people are working to verify the proof and extract its strategy, and said very early analysis suggests it draws on the full power of the Langlands program and arithmetic geometry. Tao added that moderators had since restricted that thread to Lean-related discussion. He expects the digestion to take time, noting that “at least we have a human author to communicate with” in this case.
A Passing Lean Check Settles Less About the Riemann Hypothesis Than It Appears To
A Lean kernel confirms that a formal theorem follows from the definitions written in the code. It cannot confirm that those definitions say what mathematicians mean by the Generalized Riemann Hypothesis, and a human has to read the statement and every definition it depends on to establish that. We made the same point in our coverage of the Lean proof refuting the Exponential Time Hypothesis, where the formal statement and the paper were checked against each other by people. In this case there is no paper to compare against. The Zulip effort therefore has two separate jobs: confirming the archive compiles and states the right theorem, and recovering a strategy a number theorist can follow. The BBC reported that the announcement came a day after Anthropic’s stock market debut. Alpöge called the timing horrible and wrote that the cynical reading of it is false. Neither the timing nor the machine check answers the second question, which only human reading can.
Why This Matters for Engineering Leadership
Alpöge’s release is a public example of an AI system producing an artifact that passes its mechanical check before any person understands it, including the person who directed the work. Engineering teams accepting large AI-generated changes face the same split on a smaller scale. A green build plays the role of the Lean kernel: it proves the code satisfies the tests as written, and it says nothing about whether the tests state the right requirement or whether anyone on the team can explain the design. Alpöge’s choices offer a usable pattern. He refused to publish 250,000 pages that nobody could review, he shipped the one artifact that can be checked mechanically, and he committed to producing human understanding afterwards. Tao’s ranking supplies the matching metric. Leaders can track AI-produced work by how digested it is: whether a named person has read the specification, whether someone can explain the strategy, and whether a readable write-up exists. Work that has only passed its checks belongs at the bottom of that list, however strong the result looks.
FAQ
Has the Riemann Hypothesis been proved?
Not in any accepted sense yet. Levent Alpöge announced a Lean proof on 20 November 2026, and mathematicians are still working to verify it and extract its strategy.
Who claims to have proved the Riemann Hypothesis with Claude?
Levent Alpöge, a mathematician who works at Anthropic. He says the proof was developed with Claude over roughly six weeks and also covers the Generalized Riemann Hypothesis.
Why was the Riemann Hypothesis proof released only in Lean?
Alpöge wrote that a fully detailed written version ran past 250,000 pages and would be unreadable. He chose to publish only the Lean formalization and plans a proper write-up later.
What did Terence Tao say about the Riemann Hypothesis proof?
Tao called it an enormous Lean-only proof, ranked it the least digested of three AI-assisted results, and said digesting it will take some time.
Does a Lean proof check settle whether the Riemann Hypothesis is true?
A passing check shows the formal theorem follows from the definitions in the code. People still have to confirm those definitions state the Riemann Hypothesis as mathematicians mean it.
