In October this blog looked at OpenAI’s dump of 372 math result families and asked what Lean does and does not certify. Six weeks later the same question applies to the biggest claim yet. A mathematician at Anthropic says he and Claude have proved the Riemann hypothesis. The proof was released as a Lean project, not as a paper.
What happened
On Friday, 20 November 2026, Levent Alpöge, a number theorist who works at Anthropic and is a junior fellow at Harvard’s Society of Fellows, posted a four-page note, remark.pdf, on his website. It opens: “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis.” He says the method also gives the Generalized Riemann Hypothesis and that, with more time, he expects it to reach the Extended version. The note links to the proof as fourteen archive files (lean.tar.gz.part01 to part14). Alpöge writes that he chose to publish “only Lean” because posting the machine-generated human-language version, which he says ran past 250,000 pages, would be “slop”.
The Riemann hypothesis dates from 1859. It says every non-trivial zero of the zeta function has real part 1/2, which in practice means the primes are spread out as regularly as possible. It is one of the Clay Institute’s million-dollar Millennium Prize problems.
On the timing, Alpöge writes that the final step came “late Sunday night” (16 November). He then messaged Dario Amodei about whether he could announce it. Anthropic listed on Nasdaq on 19 November, and the note went up the next day. He acknowledges “an obvious cynical view, to do with the timing” and says it is false.
What I found in the archive
I downloaded all fourteen parts. They total 239,275,554 bytes and unpack to a Lean 4 project pinned to toolchain v4.33.1 and a fixed Mathlib commit. It contains 175,187 .lean files and 17,374,215 lines, which matches the counts in the public mirrors (jsmorph/rh, Tomodovodoo/Riemann). For scale, Mathlib, the community library that hundreds of contributors have built since 2017, is on the order of two million lines.
The part a reviewer reads first is short. Main.lean is the whole interface:
theorem riemannHypothesis : RiemannHypothesis := rhOfAll
theorem riemannZeta_eq_zero_re : RiemannZetaZerosOnLine := rhOfAll
theorem LFunction_eq_zero_re : DirichletLZerosOnLine := claimAll.2.1
theorem LFunction_eq_zero_re_of_isPrimitive : PrimitiveDirichletLZerosOnLine := critAll
RiemannHypothesis is Mathlib’s own definition, not one written for this project, which matters a great deal: the classic way to “prove” a famous theorem in Lean is to prove a subtly weaker statement. The other statements are spelled out in Statements.lean in about fifteen lines. They cover the zeta zeros and the zeros of every Dirichlet L-function in the critical strip. So the “GRH” proved in the Lean code is GRH for Dirichlet L-functions, the most-used form. It is not the full generalized conjecture for every automorphic L-function.
A plain-text search of all 175,187 files found no axiom declarations, no sorry or admit, and none of the usual escape hatches (native_decide, implemented_by, @[extern], unsafe). That is encouraging but it is not proof. A grep can miss tricks hidden in macros or metaprograms. The real test is to build the project and run #print axioms riemannHypothesis, which should report only Lean’s three standard axioms. I did not compile it: 17 million lines against Mathlib needs serious hardware and many hours.
Who has checked it
Two mathematicians are on record. Andrew Sutherland of MIT told Scientific American that he had independently compiled Alpöge’s result in Lean, and called it “an historic moment”. TechCrunch updated its story to say several participants in the Lean community’s Zulip discussion had confirmed that the code compiles. Number theorist Chris Birkbeck noted there that the argument uses Galois representations attached to automorphic representations of GL_n over CM fields, base change, Jacquet–Shalika and more. In other words, the proof uses much of the heavy machinery of modern number theory.
Compiling is strong evidence of a different kind from peer review. If the kernel accepts the file and the axioms are clean, then the stated theorem follows. The remaining risks are that a definition somewhere in Mathlib does not mean what mathematicians think it means, or that the toolchain has a bug. Both are small, and both can be checked. No human, however, yet understands the argument. Alpöge says so himself: he presents “a complete argument before having mastered a massive neighbourhood around its mathematics.”
What Claude did
The note describes a division of labour. Alpöge supplied the strategy. He wanted to carry over Stepanov’s method, which proves the Riemann hypothesis for curves over finite fields, to the integers. The obstacle is that the method needs “multiplicity”, which over the integers is “something like differentiating whole numbers”. Claude contributed much of the technical work on automorphic forms for higher-rank groups, an area where Alpöge admits “little mastery”. The key step was chaining theta lifts, a way of moving automorphic forms between groups. Each lift adds one unit of multiplicity. In the note’s most quoted line, Claude proposed that for a prime sum of length X the chain should have length √X.
A timeline that, the note says, Claude produced after searching the work starts on 6 October 2026 with a “Stepanov determinant”. The iterated theta lifts arrive on 15 November and GRH on 16 November. That is about six weeks from the key idea to a fully formalized proof. Alpöge warns that parts of the timeline paraphrase a Claude summary “which might make zero sense”.
It was a long road. In August, Anthropic reported that an unreleased Claude had raised the proven share of zeta zeros on the critical line from 41.6% to 67.2%, using about 60 subagents and 31 million output tokens. Anthropic said explicitly at the time that this was not a proof. Alpöge was also behind this summer’s Claude-generated Jacobian conjecture result.
A human proof, the same day
Hours later, Stanford’s Ravi Vakil, president of the American Mathematical Society, posted a draft by Stanford postdoc Alessandro Masullo that claims GRH. It is “all human, with no AI”, and Brian Conrad has been reviewing it for months. Vakil wrote that the plan had been to polish it for a year, “now it seems best to put out the current draft.” Masullo’s draft is the reverse case: readable but not yet formalized. A Zulip thread on formalizing it has already started.
Why it matters for IT
- Verification became the bottleneck, and code solved it. Nobody can referee 250,000 pages, but a computer can check 17 million lines. This is the pattern for AI-generated software too: when output outgrows review, trust moves to machine-checkable specifications. What matters is whether the 15-line statement says what you mean. The 17 million lines below it matter much less.
- Specification review is the new code review. The most useful thing a human can do with this proof today is read
Statements.leanand Mathlib’s definitions. Developers who use AI coding agents should recognize this: reviewing the tests and the contract matters more than reading every generated line. - Capability is moving fast. In August Anthropic said Claude had not proved the hypothesis and its methods were not expected to. Three months later, an employee working with Claude claims the full result.
Open questions
- No independent understanding yet. A compiled proof with no human explanation tells us the statement is true. It does not tell us why, which is most of what mathematicians want from a proof.
- Axiom audit. A published
#print axiomsoutput from several independent builds would settle the “hidden assumption” question. The Clay Institute’s prize rules also expect publication and two years of acceptance, so a Lean-only proof is uncharted ground. - Conflict of interest. The author is an Anthropic employee, the tool is Anthropic’s product, and the timing came one day after a record IPO. Alpöge addresses this directly. The code is public precisely so that nobody has to take his word for it.
- What Claude was. The note does not say which model or how much compute was used, so the result cannot yet be reproduced as an AI-capability claim.
Sources
- Levent Alpöge, remark.pdf (20 November 2026), and the Lean archive
alpo.ge/lean.tar.gz.part01–14(downloaded and inspected for this post) - Scientific American, Anthropic mathematician claims to prove Riemann hypothesis, most famous problem in all of math
- TechCrunch, AI claimed a solution to the 167-year-old Riemann hypothesis — and a human did, too
- Ravi Vakil, Hurried announcement: Alessandro Masullo, GRH
- Anthropic, Learning more about Claude’s mathematical capabilities (10 August 2026)
- Unite.AI, Levent Alpöge’s Unverified Riemann Claim Puts Claude’s Lean Code in Focus
- Public mirrors: jsmorph/rh, Tomodovodoo/Riemann
- Clay Mathematics Institute, The Riemann Hypothesis