Ai Engineering 7 min read

An Anthropic mathematician says he and Claude proved the Riemann hypothesis. All he has released is Lean code

Levent Alpöge announced on November 20 that he and Claude have a formalized proof of the Riemann hypothesis and its generalized form, published as 14 archive files of Lean with no paper. Nobody has independently verified it. Here is what the statement says and what checking it involves.

Levent Alpöge, a number theorist who works at Anthropic, said on Friday, November 20, 2026, that he and Claude have found a proof of the Riemann hypothesis, the 1859 conjecture about the zeros of the zeta function that governs how prime numbers are distributed. In a four-page statement on his personal site, he writes that the method also settles the Generalized Riemann Hypothesis (GRH), which extends the conjecture to a wider family of L-functions, and that the argument has been formalized in the Lean proof assistant “without assumptions.”

As of this writing the claim is unverified. No independent mathematician has publicly confirmed it, Anthropic has not published anything about it, and there is no paper. Scientific American, which broke the story, describes it as a claim and says experts should be able to check fairly quickly whether the hypothesis is correctly encoded and whether the proof holds. The Riemann hypothesis is one of the Clay Mathematics Institute’s Millennium Prize problems, with a $1 million award attached, and it has a long history of announced proofs that did not survive review. This one differs in a single respect that matters: what has been released is machine-checkable.

What was actually released

The only artifact is code. The statement’s footnote points to 14 files, lean.tar.gz.part01 through part14, hosted on the same site. Alpöge says a human-readable version exists only as a generated document of more than 250,000 pages, which he chose not to post, calling that kind of output unreadable. He says he will write the argument up properly later, and that he expects it to be “massively simplified,” possibly soon.

He is also frank about the limits of his own understanding. He describes presenting a complete argument before mastering much of the mathematics around it, says he has little command of automorphic forms on higher-rank groups (the area where the late steps happen), and warns that parts of his own summary paraphrase Claude’s descriptions and may not make sense as written. The statement does not name the Claude model, the number of agents, the compute used, or the Lean and Mathlib versions.

The scope is worth stating carefully. He claims the Riemann hypothesis and GRH as done and formalized. The Extended Riemann Hypothesis, a further generalization, is something he says he expects to reach with more time, not a result.

The route he describes

The statement gives a 20-item timeline, which Alpöge says Claude compiled, running from October 6 to November 16, 2026.

The starting point is Stepanov’s method. In the 1960s Sergei Stepanov gave an elementary proof of the Riemann hypothesis for curves over finite fields (a setting where the analogous statement was first proved by André Weil) by constructing an auxiliary polynomial that vanishes to high multiplicity at many points. The obstacle to copying this over the integers is well known: there is no obvious way to differentiate, so there is no obvious notion of multiplicity. Alpöge’s account is that he and Claude found a substitute using theta lifts, a tool from the theory of automorphic forms that transfers objects from one group to another. Each lift in a chain plays the role of one extra order of vanishing.

The single line he highlights as the moment that will startle other mathematicians: for a sum over primes of length X, Claude proposed a chain of about √X successive theta lifts. The timeline lists the Stepanov-style determinant on October 6, a run of results on November 15 (iterated lifts between definite unitary groups, bounds on their norms and divisibility, a seesaw identity, and a weakened form of Vandiver’s conjecture needed to handle leftover cases), and then one entry for November 16: “GRH.”

None of this can be assessed from the statement. It is a list of ingredients, several of them deep, with no indication of how they connect. That is by design: he says the Lean is the proof and the prose is a courtesy.

Why a Lean proof still needs human checking

A proof that compiles in Lean is correct relative to three things, and each has to be inspected by people.

First, the statement. The kernel verifies that the final theorem follows from its definitions. It cannot verify that the theorem named RiemannHypothesis says what mathematicians mean by it. Mathlib already contains a standard formulation, so the first check is whether the release proves that exact statement or a private definition that only resembles it. A subtly wrong definition of the zeta function, or of a nontrivial zero, can make a false or trivial statement look like the real one.

Second, the axioms. Lean reports which axioms a theorem depends on. A clean result uses only the three standard ones. Any sorry (a placeholder for a missing proof), any added axiom, or any use of unsafe escape hatches would undercut “without assumptions,” and this is a mechanical thing to audit.

Third, the build. A project this large has to be reproduced on independent machines, against a pinned Mathlib, with nothing patched underneath. Anthropic’s earlier zeta work was run through a comparator validation tool for this reason.

Experienced formalizers can do the first two checks in hours once they have the files building. If the statement matches Mathlib’s and the axiom list is clean, the claim becomes extremely strong even with no human able to follow the argument. If either fails, it collapses just as quickly.

Reasons for caution, and for attention

The reasons for caution are plain. The base rate for Riemann hypothesis announcements is terrible. The proof was, by Alpöge’s own timeline, completed five days before the announcement. He credits a long list of people, including Peter Sarnak, Manjul Bhargava and Maksym Radziwiłł, but being thanked is not the same as having verified anything, and the statement names no reviewer. He also says he took the question of announcing to Anthropic chief executive Dario Amodei, that there was “a constraint to respect,” and that the decision to wait no longer was his alone. He acknowledges that the timing invites a cynical reading and rejects it. Anthropic began trading on Nasdaq on November 19; the statement does not say what the constraint was.

The reasons to pay attention are also real. Alpöge is a serious number theorist, not an outsider, and he reviewed the result Anthropic published in August, when a swarm of about 60 Claude agents raised the proven share of zeta zeros on the critical line from 41.6% to 67.2%, with a Lean formalization and outside review by Brian Conrey and Dan Goldston. Three days before this announcement, Anthropic published a research post describing two open problems in black hole physics that Claude agents worked on for about six weeks: a three-loop calculation showing the Hawking-Page transition stays first order at weak coupling, and a proof that one version of the modular bootstrap bounds the lightest black hole at c/9.1266 and can go no lower, with the second result formally verified in Lean and checked by a physicist. Long autonomous runs ending in formalized results are now routine output from these labs. A proof of GRH would be a different order of result, but it is the same kind of artifact.

What to watch

The questions that settle this are concrete, and most can be answered within days:

  • Does the archive build from scratch, and against which Mathlib version?
  • Is the final theorem Mathlib’s own statement of the Riemann hypothesis, and what does the GRH statement look like?
  • What does the axiom audit show?
  • Does Anthropic publish an account with the model, the compute and the degree of human steering, as it did in August and for the physics results?
  • Do specialists in automorphic forms find the theta-lift chain plausible once someone extracts a readable outline from the code?

Until the first three are answered by people outside Anthropic, the accurate description is the one in the headline: a claim, by a credible mathematician, backed by code nobody else has yet vouched for. Scientific American has said it will update its report as verification proceeds.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Engineering

How to Use Claude Across Excel and PowerPoint with Shared Context and Skills

Learn how to use Claude's shared Excel and PowerPoint context, Skills, and enterprise gateways for faster analyst workflows.

Ai Engineering

PSI says a non-public Anthropic model proved the Aaronson-Ambainis conjecture, with a full Lean check

Physical Superintelligence says its research system, running an unreleased Anthropic model, proved a central open problem in quantum query complexity by day three of a run. The 9,454-line Lean formalization uses no added axioms; independent review has only just started.

Ai Engineering

Claude Formalized Fermat's Last Theorem in Lean After 11 Days of Autonomous Work

Anthropic announced Claude produced a complete machine-checked Lean formalization of Fermat's Last Theorem after working largely autonomously for 11 days, generating verified proofs of 30,300 intermediate theorems.

Ai Engineering

OpenAI Claims Its Model Falsified the Exponential Time Hypothesis. The Verifying Starts Now.

OpenAI announced October 26 that an internal reasoning model produced a Lean-checked 2^o(n)-time algorithm for 3-SAT, which would falsify the Exponential Time Hypothesis, alongside claimed Yang-Mills and Kakeya results, with academic shepherding as the verification mechanism and a prior 722-paper release that saw three withdrawals as the reason for skepticism.

Ai Engineering

Claude Haiku 5.5 Collapses Small-Model Prices to $0.10 Per Million Input Tokens

Anthropic released Claude Haiku 5.5 on October 7 with GDPval 1620 against Haiku 4.5's 735, an adjustable effort setting, and pricing of $0.10/$0.50 per million tokens for prompts under 100k, roughly 90% cheaper than its predecessor.