Riemann Hypothesis: a proof found with Claude passes Lean verification, but no mathematician understands it yet
Mathematician Levent Alpöge, who works at Anthropic, released on November 20 a proof of the Riemann Hypothesis found with Claude, delivered only as 17 million lines of Lean code. The Lean and Mathlib communities certify that the proof is valid. No one can yet explain why it works, and the Clay prize is years away at best.

The Riemann Hypothesis, posed in 1859, states that all the “non-trivial” zeros of the zeta function have real part 1/2. It is the central question of number theory: it governs how prime numbers are distributed. On November 20, a four-page note posted by Levent Alpöge, a mathematician employed by Anthropic, claimed it has been proved. “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis,” he writes. The method, he says, extends to the Generalized Riemann Hypothesis.
What makes this announcement unusual: there is no paper. The proof is delivered only in Lean, a language in which a program, the kernel, checks every step of an argument.
What has been verified
In under two days, machine verification was carried out at scale. On November 21, the Lean and Mathlib communities, endorsed by the Lean FRO and the Mathlib Initiative, published a joint statement: they say they have “full confidence” in the proof and “certify” that both the statement and the proof are correct.
Mathematician Andrew Sutherland (MIT) lists the checks on a tracking page:
| Check | Result |
|---|---|
| Compilation of the 175,187 modules (Lean 4.33.1) | no errors |
| Axioms used by the five main theorems | Lean’s three standard axioms only |
| Statement compared with Mathlib’s by two independent tools | identical |
| Checked by two independently written kernels (nanoda, con-leche) | accepted |
| Separate offline rebuild on another machine | export identical byte for byte (61.6 GB) |
The third row is the key one. The verified statement is Mathlib’s, written and reviewed in Lean’s reference mathematical library well before this announcement. The code therefore does not prove a disguised version of the problem. The scope is precise: the Riemann Hypothesis and its generalization to Dirichlet L-functions, not automorphic L-functions in general, according to the verification page.
What no one knows yet
The proof runs to about 17 million lines. The author writes that a written version “containing sufficient detail” would exceed 250,000 pages, and that he chose not to post that “unreadable document”. He promises a proper write-up and expects quick simplifications.
In the meantime, no one can explain the argument. Terence Tao writes that the strategy appears to use “the full power” of the Langlands program and arithmetic geometry, and that digesting it “will take some time”. He refers to “an internal AI model”. Alpöge’s note describes Claude’s role in very strong terms: near the finish line, the AI “far outpaced” him. According to the timeline attached to the note, the first key step dates from October 6.
Anthropic had issued no official statement on the result as of the morning of November 22. The note itself raises the timing: it came out the day after the company’s stock market debut. The author writes that he consulted the company’s leadership before announcing, and calls a cynical reading tied to that timing “false”.
Navier-Stokes, the other announcement of the day
On the same day, Levent Alpöge and Tristan Buckmaster (NYU) posted a 79-page paper claiming finite-time blow-up for the three-dimensional Navier-Stokes equations with no external force. It is another Millennium Prize problem. Its status is very different:
| Riemann | Navier-Stokes | |
|---|---|---|
| Readable text | no | yes, 79 pages, needs a professional write-up according to Terence Tao |
| Lean formalization | about 17 million lines | about 600,000 lines |
| Independent verification | carried out, declared conclusive | incomplete: numerical certificate not released |
The repository says so itself: the formalization “contains no data”. The numerical certificate the proof depends on has not been released, nor has the interval-arithmetic computation that validates it. A November 20 patch has not been rebuilt either.
What about the million dollars?
Not for a long time. The Clay Institute’s rules set three conditions: publication in a qualifying outlet, at least two years after that publication, and general acceptance by the mathematical community. A Lean archive, even a verified one, meets none of the three today.
The announcement extends a series. In October, the start-up Axiom said it had formalized the classification of finite simple groups in Lean. In early November, a system using a non-public Anthropic model found a proof of the Aaronson-Ambainis conjecture. The jump in scale is stark: from a 2008 conjecture to the most famous problem in mathematics.
Our take
For the first time, a major problem in mathematics appears to be solved before any human understands the solution. The Lean verification is solid: the statement comes from Mathlib, several independent checkers accept the proof, and the communities involved have publicly committed to it. It is therefore reasonable to treat the result as established, pending the checks still under way. But mathematics is not just a true-or-false verdict. Until the argument is explained, it teaches researchers nothing, and that work is only starting. For AI, the signal is considerable: a model contributed decisively to a result that generations of mathematicians could not reach. One point to watch: Anthropic, which went public the day before, has said nothing official yet.
Sources
- Levent Alpöge — proof announcement note (remark.pdf, November 20, 2026) Primary source
- Lean and Mathlib communities — Lean verification of the Riemann Hypothesis (November 21, 2026) Primary source
- Andrew Sutherland (MIT) — Verification files for the Lean proof of the Riemann Hypothesis Primary source
- Terence Tao — Unforced Euler singularity, unforced Navier-Stokes singularity, and the Generalized Riemann Hypothesis (November 20, 2026) Primary source
- L. Alpöge and T. Buckmaster — Finite-time blow-up for the Navier–Stokes equations (GitHub repository) Primary source
- Clay Mathematics Institute — Rules for the Millennium Prize Problems Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
How we work: editorial charter and methodology · our verification process · corrections log. Spotted an error? Write to contact@neuralbrew.ai.


