Timeline

A Claude-assisted proof of the Riemann hypothesis passes Lean's checkers

Lean checkers accept a 17-million-line proof of Mathlib's RiemannHypothesis using three standard axioms; no human-readable proof exists, and Quanta reported nobody yet understands its steps.

  • Ideas & essays
  • Models & capabilities
  • Era-defining

On 20 November 2026 Levent Alpöge, a mathematician at Anthropic, posted a four-page statement and a Lean library of about 17 million lines, saying he had discovered, working with Claude, a proof of the Riemann hypothesis. Lean is a proof assistant, a program that checks every step of a formal argument. The next day the Lean and Mathlib communities had certified that the library proves the statement of the Riemann hypothesis already in Mathlib, Lean’s shared mathematics library, and several independently written proof checkers had accepted it.

No human-readable version of the argument exists. Alpöge chose to present only Lean, and Quanta reported that nobody yet understood the proof’s many steps. The record below separates what has been machine-checked from what has not.

I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis. In fact, the method easily extends much further, resolving the so-called Generalized Riemann Hypothesis, and I expect, with more time, the Extended Riemann Hypothesis too. Claude and I have also formalized the arguments without assumptions.

— Levent Alpöge, statement

Timeline

  • 6 October. The earliest dated step in Alpöge’s own list of how the argument was assembled. The same day, OpenAI published hundreds of unreviewed proofs, two of them of a weaker “quasi-Riemann hypothesis”.
  • Late Sunday 15 November. Alpöge writes that he messaged “Dario” about announcing the result. His list ends with an item dated 16 November, “GRH”; Quanta says he obtained the result that Monday and waited several days to share it.
  • 19 November. Anthropic lists on Nasdaq.
  • 20 November. Alpöge posts the statement and the Lean source. Mathematician Terence Tao and computer scientist Scott Aaronson post reactions. Andrew Sutherland of MIT is, per Quanta, first to compile the library, shortly before noon; Simon DeDeo of Carnegie Mellon follows hours later. Stanford’s Ravi Vakil circulates a rival human claim (below). A group signing as the AHM communications group dates its critical statement.
  • 21 November. Sutherland’s verification page is stamped 12:55 UTC; the Lean and Mathlib joint statement is stamped 13:55; Aaronson adds an update that the Lean community’s checks have passed; Scientific American updates its report.
  • 22 November. Sutherland’s page, a living document, still lists some kernel runs as running (see below).

What was verified

According to Sutherland’s page, whose checks were run with Claude Code, all 175,187 modules compile on Lean and Mathlib v4.33.1; the five headline theorems depend only on Lean’s three standard axioms (propext, Classical.choice, Quot.sound); and a replay of every module through Lean’s kernel found no rejection. Two independently written kernels, nanoda and con-leche, accept the exported proof, and lean4lean accepts every compiled module (five needed a raised step limit). The exported proof is about 61.6 GB uncompressed.

The first theorem is Mathlib’s own RiemannHypothesis. A community-written file that imports only Mathlib was compared against the proved theorem with two tools, comparator and con-leche --compare. The Lean statement adds that “dozens of people” reproduced the build and that con-leche and con-ron “come with a formal proof, relative to set theory, that they cannot accept a false proof”. Sutherland also combined the proof with a separate formalisation of Robin’s and Lagarias’s criteria, giving two inequalities about divisor sums with no mention of the zeta function.

What was not

  • Still running on 22 November. Sutherland’s page listed Lean’s default and “paranoid” kernel runs on the export, and the end-to-end lake comparator run, as in progress; its statement comparison and axiom check had already passed. The Lean statement says the export was checked by “Lean’s official kernel”; the sources do not say whether that is the same run. Scientific American wrote that the comparator run Kevin Buzzard of Imperial College London had been awaiting “has now been completed”.
  • Build provenance. Sutherland notes the build ran on cloud machines without a sandbox, and that its export has not been compared with one from a sandboxed build. The result “relies on those kernels and on Mathlib’s definitions being what they are intended to be.”
  • Scope. On Sutherland’s page “GRH” means Dirichlet L-functions, all moduli and characters. Dedekind zeta, Hecke and general automorphic L-functions “are not among the statements that were checked”, and Alpöge claims only an expectation for the Extended Riemann Hypothesis.
  • Understanding. Alpöge writes that he is presenting a complete argument “before having mastered a massive neighbourhood around its mathematics”, that parts of his descriptions paraphrase a summary he calls “slop”, and that attempts at human-readable text reached over 250,000 pages. He expects the argument to be “massively simplified”. He does not name the model; Tao writes “an internal AI model”.

Reactions

Tao, in his own post, wrote that from “very initial analysis” the proof “appears to use the full power of modern developments on the Langlands program and arithmetic geometry”, and expected digestion “will take some time”. Aaronson wrote that Lean kernel bugs have “in the past, allowed proofs of 0=1”, but that no AI result with a Lean certificate had needed retraction, so his bet was that it would stand; after the Lean statement he added “The proof is correct.” He expected mathematicians to be satisfied only when they can “explain and distill and reconstruct the ideas for themselves”.

Sutherland told Scientific American “the chance that anything is fishy here is approximately zero”. Maksym Radziwiłł of NYU told Quanta by email that Alpöge’s four-page statement contains “nonsensical statements” and that “The object of mathematics … is really the question ‘why?’ and this doesn’t answer it at all”; Quanta paraphrased his view as being that a proof in Lean code short-circuits understanding. The AHM statement, whose acronym neither page expands, said “no human being can reasonably claim to be responsible for, or even to understand, such a document”, called for democratic oversight of AI companies, and urged mathematicians to work free of dependence on AI models. It was dated before the certification.

A rival human claim

The same day, Vakil, whom Quanta identifies as president of the American Mathematical Society, posted a draft by Stanford postdoctoral scholar Alessandro Masullo claiming a proof of GRH, “all human, with no AI”, which Brian Conrad had examined for months. Quanta described a 940-page manuscript and noted an earlier Masullo attempt had not held up. About eight hours later Vakil added that several people had reported problems “identified with the assistance of different AI models”, and that he made “no claim as to the correctness of the proof strategy”. No public account of the problems was found.

Anthropic’s position and timing

Quanta wrote that Anthropic had made no official announcement; this record found none on 22 November. Alpöge wrote that “the precise timing is horrible” and that there is “an obvious cynical view, to do with the timing, but it is false”. The statement does not mention the listing; Quanta presumed that was the reference. Anthropic’s own August post on a partial result said it did not expect those techniques to prove the hypothesis.

The claim sits in a run of AI mathematics results in the record, including the exponential time hypothesis refutations, Fermat’s Last Theorem formalisation and the Aaronson–Ambainis conjecture. On the same day Alpöge, Tristan Buckmaster, a professor at NYU, and others also posted claimed blow-up proofs for fluid equations. Tao pointed to a Lean community forum thread as the place where verification and attempts to extract the strategy were being followed; Alpöge says he will try to understand the argument and write it up himself.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.