Levent Alpöge, a mathematician at Anthropic, announced on Friday, November 20 that he believes he and Claude have proved the Riemann hypothesis — the 167-year-old conjecture about the zeros of the zeta function and one of the seven Clay Millennium Prize Problems. He says the proof also covers the generalized Riemann hypothesis. Rather than a paper, he released a four-page note on his website and a formal proof written in the Lean proof assistant, according to IBM Think and DataCamp.
Why it matters: if it holds up, it would be the most important mathematical result in decades and the clearest case yet of an AI system doing frontier research. Oxford's James Maynard, a Fields medalist, told IBM it would be a "historic milestone in mathematics" if it checks out. But the form of the claim — a machine-checkable proof that, by its author's own account, humans currently can't read — is itself new, and it changes what "verified" means.
What was released
| Item | Detail | Source |
|---|---|---|
| Announcement | Four-page note, "Remark," on Alpöge's site | IBM, CASRAI |
| Formal proof | 14-part archive of 175,187 Lean 4 files | CASRAI; DataCamp confirms 14 parts |
| Size | 239.3 MB compressed, over 61.5 GB unpacked | CASRAI |
| Claimed results | Riemann hypothesis and the generalized RH for Dirichlet L-functions | IBM, CASRAI, DataCamp |
| Paper or preprint | None; no journal submission | DataCamp |
DataCamp reports a footnote in which Alpöge says a natural-language version ran past 250,000 pages and that he chose not to publish it. IBM says he acknowledged the argument is extremely long and hard to read.
What Claude did
Accounts differ in detail. DataCamp says Alpöge has worked with Claude since October 6, that key steps date to November 15, and that he directed the approach — but that near the end Claude moved faster than he could follow, and he concedes limited command of one area the proof relies on. CASRAI, citing other outlets, says Alpöge set the overall architecture while Claude handled algebraic manipulations, trace formulas and technical lemmas. Neither source names the Claude model version. Anthropic had not commented as of DataCamp's report; CASRAI cites Quartz saying the company reviewed internal publication protocols before the announcement was cleared.
What has actually been checked
This is where care is needed. A Lean proof can be checked by a computer, which is why the format is attractive. But two separate questions remain:
- Does the code compile and pass the kernel? DataCamp, writing on November 20, said nobody had yet publicly reported building the files. CASRAI, a day later, reports results it attributes to MIT's Andrew Sutherland: all 175,187 modules compiled, independent kernel re-checks accepted them, and the headline theorems rely only on Lean's three standard axioms — which would imply no unfinished
sorryplaceholders. We have not seen Sutherland's results directly, so treat these as reported. - Does the formal statement mean what it should? Even a perfectly checked proof only proves the statement as written. Mathematicians still need to confirm that the definitions used — CASRAI says the main theorem is stated against Mathlib's definition of the zeta function — faithfully capture the real conjecture and aren't vacuous. CASRAI says the community treats the result as unverified pending that review.
Youness Lamzouri of the University of Lorraine told IBM that understanding the proof could take months or years. Maynard told IBM he suspects it will prove correct, but that checking it will take time.
The prize question
The Clay prize is $1 million. DataCamp and CASRAI note the rules require publication in a qualifying peer-reviewed journal followed by two years of acceptance before the institute deliberates; DataCamp estimates 2028 at the earliest. There is no conventional paper yet. The Clay Institute had not commented, per DataCamp.
What to watch
- Independent builds of the archive by other formalization groups, and published audits of the formal statement.
- A human-readable account of the core idea — CASRAI says it extends Stepanov's method using theta lifts — that number theorists can evaluate.
- An official Anthropic statement, and whether the Claude model involved is named.
Until then, the honest summary is: a serious claim, an unusual and partly machine-checked artifact, and no human verification of the mathematics.
Sources
- IBM Think — Anthropic mathematician claims AI helped solve a 167-year-old math mystery ibm.com
- DataCamp — Anthropic Researcher Claims AI Solves the Riemann Hypothesis, the Most Famous Problem in Math datacamp.com
- CASRAI — Anthropic Mathematician Releases 175,000-File Lean Archive Claiming Riemann Hypothesis Proof casrai.org