Levent Alpöge, a mathematician at Anthropic, says he and Claude have found a proof of the Riemann hypothesis, and he has posted it only as machine-checkable Lean code, in a four-page note on his website published on 20 November. If the code holds up, the most famous open problem in mathematics would be settled with an AI model doing much of the work. Nobody outside the project has confirmed that yet, so we downloaded the full archive and looked at what it actually contains.
What Alpöge released, and what he did not
The note is short and personal. “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis,” Alpöge writes. He says the method also covers the Generalized Riemann Hypothesis for Dirichlet L-functions, and that he expects the Extended version to follow “with more time”.
There is no paper. Alpöge says a full written version ran past 250,000 pages and calls posting it an “unreadable travesty” he wanted to avoid, so the only evidence is a Lean 4 project split into 14 downloadable parts. He promises a proper write-up and expects the argument to get much shorter. On the division of labour he is blunt: “Claude far outpaced me near the finish line.”
Inside the 239 MB archive: our check
We joined the 14 parts (239,275,554 bytes), confirmed that the SHA-256 fingerprint matches an independent public mirror on GitHub, unpacked it and scanned every file. The numbers:
- 175,187 Lean files with about 17.4 million lines, of which 141,928 files sit in a single folder named Development.
- Roughly 779,000 theorem and 34,000 lemma declarations, plus about 212,000 definitions.
- A pinned toolchain: Lean 4.33.1 and a fixed Mathlib commit, so anyone can try to rebuild exactly this state.
- The top-level file is 12 lines long. Its key line, theorem riemannHypothesis : RiemannHypothesis := rhOfAll, targets the RiemannHypothesis definition that already lives in Mathlib, the shared Lean maths library, not a definition written for this project.
- A text search found zero occurrences of sorry, admit, axiom declarations, native_decide, implemented_by, unsafe or opaque, the usual ways to smuggle an unproven step or a trust shortcut into a Lean proof.
That last point needs a caveat. A text scan is not a compile. Only a full build of the project plus a #print axioms check on the headline theorem shows whether the proof really goes through on Lean’s standard axioms. With this many files, that build takes serious hardware and time, and as of Friday evening we found no public report of a completed independent run.
From 67.2 percent to the whole conjecture
This is not Anthropic’s first brush with the problem. In August, the company reported that an unreleased Claude model had raised the proven share of zeta zeros on the critical line from 41.6 to 67.2 percent, a result we covered in Claude Advances Riemann Zeta Bound with Multi-Agent AI. Alpöge was one of the two Anthropic mathematicians who checked that paper. Anthropic wrote at the time that it did not expect those techniques to lead to a proof of the full hypothesis.
The new route is different: the note describes a descendant of Stepanov’s method, with chains of theta lifts standing in for the multiplicity trick that works over function fields. Our count of the timeline in the note gives a sense of the pace. It lists 20 steps between 6 October and 16 November, a span of 41 days. Nine of the 20 are dated 15 or 16 November, the Generalized Riemann Hypothesis is the final entry on 16 November, and nothing is dated between 20 October and 15 November.
The timing question
The note appeared one day after Anthropic’s Nasdaq debut, and on the same day Anthropic’s newsroom announced the closing of the IPO, with nothing about the proof. Alpöge raises this himself: “There is an obvious cynical view, to do with the timing, but it is false.” He writes that he messaged “Dario” (Anthropic’s chief executive is Dario Amodei) late on a Sunday night and that “there was a constraint to respect” before he could announce.
Why Lean helps here, and where it stops
Formal proofs only certify the statement that was formalized. That gap was the central issue in OpenAI’s Navier–Stokes claim. In this case the risk is smaller, because the headline theorem points at Mathlib’s existing definition of the hypothesis. The open question is mechanical: does the whole project compile, and does it rely on anything beyond Lean’s three standard axioms?
For mathematicians and research teams, the practical step is clear: reproduce the build with the pinned toolchain before citing any consequence of the result. The many published results that assume the Generalized Riemann Hypothesis can only start to be treated as unconditional once that check passes, and fully only after experts have read the argument.
Still open
- A completed, independent build and axiom check. We will add the result here when one is published.
- A human-readable proof, which Alpöge says he intends to write.
- The Millennium Prize. Under the Clay Mathematics Institute rules, a solution must appear in a qualifying outlet, at least two years must pass after publication, and it needs general acceptance in the mathematics community, so a prize decision is years away even if the code checks out.






