Blog · AI and math

Claude and the Riemann hypothesis: what was claimed, and what has been checked

On Nov 20, 2026, Anthropic mathematician Levent Alpöge said that, working with Claude, he believes he found a proof of the Riemann hypothesis, released as a Lean-only formalization; it has not been peer-reviewed and checking is ongoing. He says the same method extends to the Generalized Riemann Hypothesis, and the Lean code targets the Riemann hypothesis plus GRH for Dirichlet L-functions. According to Scientific American, MIT's Andrew Sutherland compiled the code independently, but experts still ask whether its statements match the real theorems. The same day, Ravi Vakil released a separate, all-human 940-page draft by Alessandro Masullo. This page sets out what was claimed, what has been checked so far, and how to tell whether it holds. Updated November 21, 2026.

What was claimed

Alpöge's four-page statement opens: "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis." He writes that the method extends further, to the so-called Generalized Riemann Hypothesis (GRH), and that he expects the Extended Riemann Hypothesis only with more time. He gives no human-readable proof; the proof he points to is a Lean archive in 14 parts on his website.

  • The Riemann hypothesis (RH) says every nontrivial zero of the Riemann zeta function has real part 1/2 (Mathlib's definition). GRH asks the same of a wider class of L-functions, and Scientific American's analysis notes the claim is technically for GRH.
  • What the Lean code states: according to the README of a public mirror, the main file states as theorems RH as defined in Mathlib (Lean's math library) and that the zeros of every Dirichlet L-function in the critical strip lie on real part 1/2.
  • What it doesn't state: per the same mirror README, the targets are RH and GRH for Dirichlet L-functions, not GRH for wider families such as Hecke or automorphic L-functions. The Extended Riemann Hypothesis is only expected, not claimed.
  • The statement names only Claude, with no model version, tool, line count or build time.
  • When we checked Anthropic's news page on November 21, it had posted no statement about the claim. Anthropic's August 2026 research on zeta zeros is a separate, earlier result.

What Claude did, according to Alpöge, Tao and Aaronson

  • Alpöge writes that Claude moved ahead of him near the end. His example: Claude suggested a chain of theta lifts of length √X for a von Mangoldt sum of length X.
  • He says he guided the work to the right setting, and that he has little mastery of automorphic forms on higher-rank groups.
  • He says he and Claude also formalized the arguments in Lean without assumptions, and that the formalization is finished.
  • A generated human-language write-up passed 250,000 pages, he says, and Claude called it the GRH book. He chose to present only Lean rather than post that text, and he plans to write the argument up properly once it has been simplified.
  • The statement includes a rough timeline that Claude provided, dated October 6 to November 16, 2026, with GRH as the last item; he warns that parts of it may not make sense.
  • Terence Tao describes it as an enormous Lean-only proof generated by Alpöge using an internal AI model. Scott Aaronson writes that Alpöge did the work with Anthropic models.

What a Lean formalization means

Lean is a proof assistant: a program whose kernel checks every step of a proof written in its language. Alpöge chose to release only Lean, so for now the checking is done on the code rather than on a paper. Two questions matter: does the code compile with no hidden assumptions, and do its definitions say what mathematicians mean by RH and GRH. Tao points readers to a thread on the Lean Zulip chat to follow verification.

The code is large. The maintainer of a public mirror counts about 17.4 million lines in 175,187 Lean files, a figure cited by Scott Aaronson. Alpöge's statement gives no line count.

What has been checked so far

Status as of November 21, 2026, about 01:25 UTC. Each row names who reported it; none of these checks was rerun for this page.

CheckStatusReported by
Lean code publishedYes, 14 archive parts on alpo.geAlpöge's statement
No sorry or custom axiomsA text scan found 0 sorry, 0 admit and 0 axiom declarations. The README says a scan is not a substitute for a build.Mirror maintainer
Compiles in LeanAndrew Sutherland (MIT) compiled it without issues, and the Lean kernel rechecked it using only the three standard axiomsAccording to Scientific American
Lean statements match the real theoremsOpen question, raised by Kevin BuzzardScientific American
Comparator run (redefined symbols, several kernels)Had not yet succeeded at publicationScientific American
Human-readable proofNone yet; Alpöge plans to write oneAlpöge's statement
Peer reviewNone-

The separate human proof attempt by Alessandro Masullo

On the same day, Stanford's Ravi Vakil posted that Alessandro Masullo has a proposed proof of GRH, and that Brian Conrad has been looking closely at it for months. Vakil writes: "It is all human, with no AI." Scientific American describes Masullo as "a young mathematician, also at Stanford".

  • The draft, The Music of the Primes, runs to 940 pages and aims at GRH for Hecke L-functions of any number field, a wider family than Dirichlet L-functions.
  • It is unfinished in places: in the PDF, some sections, including 13.1 and 13.2, are empty, a theorem in Chapter 13 carries the footnote "Insert statement.", and one reference is unresolved.
  • No Lean formalization comes with it. Vakil says the plan had been to bring Masullo to Stanford for a year to polish the writing, and asks people to check it, formalize it and fill any small gaps.
  • Brian Conrad told Scientific American: "It's undergoing major revision."

What experts say

  • Terence Tao (UCLA) writes that from a very initial analysis the strategy appears to use the full power of the Langlands program, and: "I imagine the digestion process will take some time."
  • Scott Aaronson (UT Austin) writes: "my bet is that the proof will stand", and he sees no indication so far of Lean exploits.
  • Andrew Sutherland (MIT), according to Scientific American: "Honestly the chance that anything is fishy here is approximately zero."
  • Kevin Buzzard (Imperial College London) asks, per Scientific American: "does this correspond to the actual theorem statement".
  • Coverage differs in tone: one Scientific American article says no one has verified the proof and some mathematicians are skeptical, while another says independent mathematicians attest it is almost certainly true. Both are dated November 20.

What happens next, and how to tell if it holds

  • A successful comparator run: Buzzard was still waiting for one, per Scientific American.
  • An answer to whether the Lean definitions match the real RH and GRH, the question Buzzard raises.
  • A readable proof: Alpöge says the argument is very long for now and he expects it to be greatly simplified. Aaronson expects that distilling its ideas will probably take years.
  • Peer review: none has happened yet for either Alpöge's code or Masullo's draft.
  • For Masullo's draft: Conrad's revision, and outside checking and formalization, which Vakil invites.

Why it would matter if it holds

  • For number theory: Hector Pasten told Scientific American that, if right and understood, it could be "something like a Rosetta Stone." Carlo Pagano told the same magazine it could open "brand new avenues of thinking in number theory."
  • For method: Tao notes the strategy appears to draw on the full power of modern work on the Langlands program.
  • For security: Scientific American's analysis says GRH gives no shortcut against RSA encryption.
With Crewmark

Agents for everyday work

Crewmark turns coding agent sessions into AI employees that take the next issue from a board on their own. Claude Code and Codex employees run natively; other agents, such as OpenCode, DeepSeek Harness and Gemini, connect through ACP (acpx).

FAQ

Was the Riemann hypothesis proved?

Not confirmed. Levent Alpöge claims a proof made with Claude, released only as Lean code. An independent compile is reported according to Scientific American, but whether the Lean statements match the real theorem is still open, a comparator run was pending, and there is no peer review yet.

What did Claude do?

Alpöge says he worked with Claude, that Claude suggested steps such as a chain of theta lifts, and that the two formalized the argument in Lean. He says he guided the work. His statement names no Claude model.

What is Lean?

A proof assistant whose kernel checks each step of a formal proof. Alpöge released only Lean code; a public mirror's maintainer counts about 17.4 million lines, a figure Scott Aaronson cites.

Who is Levent Alpöge?

A mathematician at Anthropic, per Scientific American. Terence Tao's post also reports AI-assisted, Lean-formalized blowup results for the 3D Euler and Navier-Stokes equations with Buckmaster and Coiculescu.

Is there a human proof?

A separate one is proposed. Ravi Vakil released a 940-page, all-human draft by Alessandro Masullo aiming at GRH for Hecke L-functions. It is unfinished in places, has no Lean formalization, and Brian Conrad told Scientific American it is undergoing major revision.

What did Terence Tao say?

Tao wrote that Alpöge, using an internal AI model, generated an enormous Lean-only proof, that the strategy appears to use the full power of the Langlands program, and that digesting it will take some time.

Checked against Levent Alpöge's statement, the public Lean mirror's README, Terence Tao's and Scott Aaronson's posts, Ravi Vakil's announcement, Masullo's draft, four Scientific American articles and Anthropic's news page on November 21, 2026.

Give your agents a standing job.

Crewmark turns coding agent sessions into AI employees with roles, goals, working hours and a heartbeat.

Crewmark is open source: star us on GitHub