Claude-assisted proof of the Riemann hypothesis passes Lean verification

The Lean and Mathlib communities said on November 21, 2026 that they certify a formal proof of the Riemann hypothesis released a day earlier by Levent Alpöge, a mathematician at Anthropic who says he found the argument while working with Claude. In a joint statement endorsed by the Lean FRO and the Mathlib Initiative, the communities wrote that "both the statement and the proof of the Riemann Hypothesis are correct" and that they have "full confidence" in the result, which arrives as a Lean library of about 17 million lines of code that depends directly only on Mathlib.
Alpöge published the work on November 20 as a short note on his personal website together with the full Lean source. The note opens with the sentence "I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis" and claims that the same method extends to the Generalized Riemann Hypothesis. The formal release compiles into 175,187 Lean modules, ships as a 239 MB compressed source archive split into 14 parts, and produces a proof export of roughly 61.6 GB once unpacked, according to the verification page maintained by Andrew Sutherland at MIT, who ran the first public checks with Claude Code.
The Riemann hypothesis, stated by Bernhard Riemann in 1859, says that every nontrivial zero of the Riemann zeta function has real part one half. It is one of the seven Millennium Prize Problems of the Clay Mathematics Institute and governs how regularly the prime numbers are distributed. The Lean statement is the first institutional endorsement the claim has received from a formal mathematics organization.
··········
WHAT THE LEAN AND MATHLIB COMMUNITIES CERTIFIED ABOUT THE RIEMANN HYPOTHESIS PROOF.
The joint statement describes four checks, adapted from criteria published by the Xena Project. The first concerns the statement itself: the theorem being proved is Mathlib's own `RiemannHypothesis` definition, which the communities describe as carefully vetted by experts in Lean and in number theory. A formal proof is only as meaningful as the formal statement it proves, and using an existing library definition removes the risk that the release quietly proves a weaker or differently worded claim.
The second check is compilation. The statement says dozens of people compiled the library independently, reproduced the same raw proof export file, and reported no errors or build failures. The third check confirms that the proved theorem matches the Mathlib statement: a community file isolates the statement while importing only Mathlib, and two tools, `comparator` and `con-leche --compare`, confirmed the match.
The fourth check is the validity of the proof. The exported proof was accepted by Lean's own kernel, by the independent checkers `con-leche` and `con-ron`, and by several third-party checkers written in other programming languages. The communities add that `con-leche` and `con-ron` carry a formal proof, relative to set theory, that they cannot accept a false proof, and that every check ran on "an inert export of the proof, out of reach of hypothetical malicious code in the Lean proof sources." That precaution addresses a specific risk of very large formal developments, where custom tactics or metaprograms could in principle tamper with the environment during compilation.
··········
HOW THE CLAUDE RIEMANN HYPOTHESIS PROOF WAS CHECKED: TOOLS AND RESULTS.
Sutherland's MIT page records the individual checks and their status as of its last update on November 20 at 22:05 UTC. The five headline theorems, the Riemann hypothesis plus four related statements, depend only on Lean's three standard axioms, `propext`, `Classical.choice` and `Quot.sound`. The page states that "GRH" in the release means the analogous statement for Dirichlet L-functions only, and the community certification addresses the Riemann hypothesis as defined in Mathlib.
........
Check | Tool | Reported result |
Compilation | Lean v4.33.1 with Mathlib | Zero errors across all modules |
Kernel replay | leanchecker | 175,187 of 175,187 modules accepted |
Independent kernel | nanoda (Rust) | Full export accepted, about 28 minutes on 80 threads |
Independent kernel | con-leche, verified mode | 1,468,418 declarations accepted, about 3 minutes on 64 threads |
Statement comparison | con-leche --compare | All five theorems justified |
Independent kernel | lean4lean | 175,187 of 175,187 accepted |
Full comparison | lake comparator | Still running at last update |
........
The page is candid about the limits of these runs. The build used cloud machines without a sandbox and compiled each file by calling `lean` directly, and only a 77-module sample was compared against a standard `lake build`. The export carries the same provenance, and its checksum had not yet been compared with a sandboxed export. Sutherland also notes that a passing check shows the formal statements follow from the three standard axioms according to the kernels used, and that the result relies on those kernels and on Mathlib's definitions meaning what they are intended to mean. The community certification published the following day reports that Lean's kernel, con-leche, con-ron and several third-party checkers accepted the inert export.
··········
WHAT ALPÖGE SAYS CLAUDE DID IN THE RIEMANN HYPOTHESIS PROOF.
Alpöge's note is untitled, unsigned in its body and only a few pages long. It refers to the model simply as "Claude" and gives no version number, compute budget or running time. Its timeline of steps runs from October 6 to November 16, 2026, which places the main work in a window of about six weeks. The note describes Claude as doing much of the technical work and says that parts of its own summary are paraphrased from Claude's output; it also mentions more than 250,000 pages of material produced along the way. Its acknowledgments thank individual mathematicians and do not mention Anthropic by name.
The mathematical outline in the note starts from Stepanov's method, the polynomial technique used to prove the Riemann hypothesis for curves over finite fields, which relies on constructing auxiliary functions that vanish to high multiplicity. Because multiplicity has no direct analogue over the integers, the argument builds one from chains of theta lifts between totally definite unitary groups, with the chain length set to about the square root of the length of the prime sum being estimated. The outline then combines a Guinand-Weil explicit formula, period identities for automorphic forms, an Iwasawa tower with divisibility bounds, repeated applications of Parseval's identity together with a relative trace formula, and a weakened form of Vandiver's conjecture for class groups.
None of these steps has yet been explained in a form that number theorists can read as a conventional paper. The Lean statement itself says that activity in the mathematical community will be needed to explain the new ideas in the proof. A formal certificate establishes that the theorem follows from the axioms; it does not by itself convey why the method works or whether it yields new tools for other problems.
··········
CLAUDE AND THE RIEMANN ZETA FUNCTION: FROM 67 PERCENT OF ZEROS TO A FULL PROOF.
The November release follows a narrower result that appeared in August. A paper by Alpöge and Ralph Furman dated August 21, 2026, proved unconditionally that at least two thirds of the nontrivial zeros of the zeta function are simple and lie on the critical line, rising to 67.25% with an additional technical window, and that at least five sixths of the zeros are distinct. The paper states that "the mathematical argument in this paper was discovered and written by Claude," with the human authors verifying the proof and taking responsibility for it. Anthropic described that work at the time as an attempt by an unreleased research version of Claude on the Riemann hypothesis that fell short of the full problem but improved a related bound.
........
Date | Result | Formal verification | Role of Claude as stated |
2020 record | 5/12 (about 41.7%) of zeros simple and on the line | None | None |
August 21, 2026 | At least 2/3 (67.25% with refinement) simple and on the line | Lean 4 | Argument discovered and written by Claude |
November 20, 2026 | All nontrivial zeros on the line (Riemann hypothesis) | Lean 4, certified November 21 | Claude did much of the technical work |
........
The jump from a proportion of zeros to all of them is the whole difference between a strong partial result and the hypothesis itself. Earlier this year Claude also completed a Lean formalization of Fermat's Last Theorem, covered in our report on Claude's 11-day Fermat's Last Theorem formalization, and OpenAI presented a proposed solution to the Navier-Stokes problem in September, described in our coverage of OpenAI's Navier-Stokes proposal.
··········
WHAT THE RIEMANN HYPOTHESIS RESULT MEANS FOR CLAUDE USERS AND RESEARCHERS.
For people who use Claude, the release does not change any product, plan or price. Anthropic has not said which model or internal configuration Alpöge worked with, whether it is available in Claude apps or the API, or how much compute the work consumed, and the August paper explicitly referred to an unreleased research version. Readers should treat the result as evidence of what Anthropic's research systems can do under expert direction over several weeks, and as distinct from what a Claude Pro or Claude Max subscription delivers in a chat session today.
Researchers who want to reproduce the verification can follow the MIT page, which lists the archive parts and recommends about 100 GB of RAM and 70 GB of free disk for a full check, or about 50 GB of RAM for the con-leche route. The checks themselves illustrate a practical workflow: Sutherland ran them with Claude Code, and the independent kernels took between roughly three and 28 minutes on many-core machines once the export existed. The heavy cost lies in compiling 175,187 modules and producing a 61.6 GB export, which is out of reach for a laptop.
For the mathematics profession the certification creates an unusual situation, a result that formal tools accept before human experts understand it. The Clay Mathematics Institute's rules for its $1 million prize require publication in a qualifying outlet and general acceptance by the mathematical community after at least two years, so a prize decision is far off whatever the formal status. Anthropic itself began trading on Nasdaq on November 19, a sequence described in our coverage of Anthropic's IPO pricing at $83 per share, and it had not issued a formal statement on the proof when the Lean certification appeared.
··········
CLAUDE RIEMANN HYPOTHESIS PROOF: CONFIRMED FACTS AND OPEN QUESTIONS.
Several facts are established by primary sources. A Lean library of about 17 million lines and 175,187 modules proves Mathlib's definition of the Riemann hypothesis using only the three standard axioms; the Lean and Mathlib communities, with the Lean FRO and the Mathlib Initiative, certified the statement and the proof on November 21 after independent kernels accepted an inert export; and Alpöge attributes much of the technical work to Claude over a timeline running from early October to mid-November.
Other points remain open. The release's generalized statement covers Dirichlet L-functions only, the model version and compute behind the work are undisclosed, Anthropic has not commented formally, and no human-readable account of the theta-lift construction exists yet. The formal certificate settles the logical question as far as Lean's kernels and Mathlib's definitions can settle it. Mathematical acceptance will depend on number theorists extracting the ideas from the code and explaining them in a form that can be taught, extended and applied to the problems that the Riemann hypothesis has been holding back for 167 years.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




