Research

Anthropic mathematician says he and Claude found a Riemann proof

Anthropic mathematician Levent Alpöge says he and Claude proved the Riemann and generalized Riemann hypotheses, releasing 175,187 Lean files; review is pending.

A plot of the Riemann zeta function on the real line, with a vertical asymptote on each side of the origin

On November 20, Levent Alpöge, a mathematician at Anthropic, published a four-page note saying he and Claude had found a proof of the Riemann hypothesis, and that the method extends to the generalized Riemann hypothesis (GRH). He released no paper, only 14 compressed archives containing 175,187 Lean files. As of the morning of November 21, no independent team had published a full compile of the code.

Key points

  • The claim is “I believe”, not “it is proved”. The note opens with “I believe I have, working with Claude, discovered a proof”; it also says the argument may contain strange suboptimal choices and should be simplified quickly. No institution had confirmed the proof as of November 21.
  • The size is checkable. The third-party mirror holds 175,187 .lean files and 17,374,215 lines of text (my count after cloning on November 21). Main.lean lists five theorems; the Riemann hypothesis itself uses Mathlib’s own definition, while the generalized case is written as 16 lines of local definitions.
  • Type-checking is not proof of the conjecture. The code contains no sorry, axiom or native_decide (I checked with grep), but no independent build log exists, and the Clay Prize requires formal publication plus a two-year wait.

Background

Bernhard Riemann posed the Riemann hypothesis in 1859: every nontrivial zero of the zeta function should have real part 1/2. It made David Hilbert’s 1900 list of problems and became one of the Clay Mathematics Institute’s seven Millennium Problems in 2000, worth $1 million. In 167 years no announced proof has been accepted by the field; the last loud claim was in 2018.

AI entered this corner of number theory only recently. On October 6, OpenAI published 722 math manuscripts produced by an internal model, and we wrote about what was in them. Later in October a Shanghai AI Lab team said its Intern-S2 model had helped with a proof of the three-dimensional Ghys conjecture, with a Lean formalization from Peking University. Alpöge’s own timeline starts on October 6, and his note thanks OpenAI for two arguments tucked quietly into that release.

Anthropic had been here before. In August, according to IT Home, the company reported that Claude had raised a lower bound on the proportion of zeta zeros satisfying the hypothesis from 41.6% to 67.2%, still well short of a proof. Three months later the same problem returned as a headline.

The facts

The note is four pages; the PDF’s metadata records 2026-11-20 12:52 UTC. The opening paragraphs give the claim, followed by a sketch of the method and a 20-step timeline. He explains why he published only Lean: laying the argument out as a document would produce something “massive” and “unreadable”, and posting that would be inappropriate, so he chose to “only present Lean”. He also says the argument is formalized “without assumptions”, a claim that checking the axioms is meant to settle.

The 250,000-page figure circulating in Chinese coverage comes from a footnote on page two. It is the author mocking himself: he did generate that document, “of course amused myself, and hit 250,000+ pages”. It is an auto-generated draft he chose not to release, not a written paper. Reading it as “a paper too long to publish” overstates how much readable material exists.

The code ships as 14 archives, lean.tar.gz.part01 through part14. A Huxiu writer downloaded them and measured roughly 229 MB compressed and about 1.4 GB unpacked. I cloned the mirror on November 21: all 175,187 files under src/ are .lean, and wc -l totals 17,374,215 lines. The mirror’s README says about 17.4 million, which matches. The toolchain is Lean v4.33.1 with Mathlib pinned at 0df444a3.

The timeline has 20 steps. The first is dated October 6 (a Stepanov determinant); the last, dated November 16, is labelled GRH. Eight steps fall on November 15 alone, including the choice to set the length of an iterated theta lift to √X — the note credits Claude with that line. The author also says the timeline is his paraphrase of a summary Claude produced and “might make zero sense”.

ItemNumberBasis
Note4 pagesalpo.ge/remark.pdf, 2026-11-20
Archives14, about 229 MBHuxiu writer’s download; about 1.4 GB unpacked
Lean files175,187Mirror repo, counted 2026-11-21
Lines of text17,374,215Same, includes blanks and comments
Timeline20 steps, 41 days2026-10-06 to 11-16
Unreleased draft250,000+ pagesAuthor’s footnote, not a paper
Lean versionv4.33.1Repo lean-toolchain

What others say

New Scientist reported the claim on November 20 and confirmed the author’s identity and affiliation. It also noted the timing: Anthropic had listed on Nasdaq the day before. A mathematician at Imperial College London gave the most practical response: “Let’s see if it compiles first.” Another mathematician, speaking to New Scientist, called the recent pace of automated proof unsettling and exciting at once.

Terence Tao wrote a preliminary inventory on his blog. On a very first read, he said, the proof appears to draw on the Langlands program and modern arithmetic geometry; the Lean Zulip thread is the place to follow verification and to extract the strategy, and “the digestion process will take some time”. Commenters asked whether anyone had checked that the proof stays inside ZFC, and one said plainly that he had not built the archive or checked its proofs.

The accounts do not line up on every point. The blog describes a Lean-only proof of GRH, while the author’s note leads with the Riemann hypothesis and then extends; the blog says “an internal AI model” without naming Anthropic or Claude, while New Scientist and the note itself point to Claude. One mathematician said that after an initial check the formalized statement “is in fact GRH”, while putting the chance of a soundness bug in Lean itself below 1%. A number theorist at the University of Manchester said Lean community experts were checking the code in parallel.

What Lean can and cannot prove

This is the layer most coverage skips. A successful type-check says the proposition written in the code has a valid proof term; it says nothing about whether that proposition is the Riemann hypothesis a mathematician means. Here the headline statement helps: it is Mathlib’s definition, external to the author’s repository, so anyone can compare it. The generalized case is weaker ground: the three GRH statements plus a zero-form statement of the Riemann hypothesis are 16 lines of local definitions in Statements.lean, and their meaning has to be read by a person.

The second layer is the toolchain. Past Lean 4 kernels have contained bugs that let invalid proofs through, which IT Home noted. A proper check therefore includes replaying the compiled environment with lean4checker and printing #print axioms to see that only standard axioms are used. The mirror’s README recommends exactly those steps. As of November 21, no public record showed anyone finishing them.

The third layer is the informal argument. According to a Science report in October 2026 cited by IT Home, an earlier OpenAI formalization of a Navier–Stokes result weakened the English statement it claimed to prove. Whether that repeats here depends on someone reading those 16 lines closely, and on that person being willing to put a name to the verdict.

Our take

For teams working on AI for mathematics, the interesting part is not “AI solved another famous problem”. It is that the deliverable changed: not a paper but a codebase that machines can check and audit. Three practices carry over — treat the formalization as the artifact, run a static scan before trusting it (sorry, axiom, unsafe), and decide in advance who verifies that the formal statement matches the prose claim.

The honest problems are just as clear. Human readability is near zero: the author says he will spend the coming period understanding and simplifying the argument, with no timetable. Verification is expensive: 170,000 files and 17 million lines need hours of compilation, a Mathlib cache and a large machine. And the news cycle runs far ahead of the mathematical one; the claim reached Chinese social media the same day, while review may take months.

Who should care: the formalization community, which may soon receive submissions of this size; science writers, who need to keep “it compiles” and “it is proved” in separate sentences; and Anthropic, which has to choose a position between a commercial story and academic caution. For developers the lesson is narrower but useful — judge an AI science claim by what it shipped, then by who is checking it.

What to watch

Four things: whether the Lean Zulip thread produces a full build log and axiom output; whether the author’s promised write-up gives a readable skeleton of the proof; the Clay path, which still runs only through publication in a qualifying outlet, a two-year wait and general acceptance by the field, with personal sites and tarballs outside the process; and whether Anthropic adopts the result officially. A reproducible build log from the community would be the first real signal.