Mathematician Levent Alpoge released a 175,187-file formal proof of the Riemann Hypothesis on November 20, written in the Lean 4 proof assistant with help from Anthropic’s Claude. The Lean Focused Research Organisation and the Mathlib Initiative said a day later that the code compiles, matches the stated theorem and passes every independent proof checker they ran.
The Riemann Hypothesis, posed by Bernhard Riemann in 1859, is one of the seven Clay Mathematics Institute Millennium Prize problems, each carrying a $1 million award. Alpoge, who holds a position on Anthropic’s technical staff and is a junior fellow at the Harvard Society of Fellows, said he reached the final steps of the proof late on November 15 after roughly 41 days of work with Claude, then consulted Anthropic chief executive Dario Amodei and mathematical colleagues before releasing it.
Because he judged a human-readable write-up would run past 250,000 pages, Alpoge published the argument only as Lean code, in a 14-part archive that expands to more than 61 gigabytes, alongside a four-page cover note. The note says the proof also covers the Generalized Riemann Hypothesis for Dirichlet L-functions, and builds on a method mathematician Sergei Stepanov originally developed for a finite-field version of the problem.
What was actually checked
Three separate things were verified, not the mathematics itself. Mathematician Kevin Buzzard’s four-point checklist, which the Lean and Mathlib communities followed, asks whether the code compiles, whether its final theorem matches the standard definition of the Riemann Hypothesis already sitting in Mathlib’s library, and whether an independent kernel accepts the exported proof.
MIT mathematician Andrew Sutherland coordinated that work and posts live results on his own tracking page. As of Sunday afternoon, it showed every one of the 175,187 compiled modules passing Lean’s kernel and a separate checker called lean4lean, with the five headline theorems resting on only the three standard axioms Lean proofs are allowed to use. A Rust-based checker called nanoda independently confirmed the exported proof in under 28 minutes. One module produced a non-deterministic hash on repeat runs, which Sutherland’s page says does not affect the result, and a slower, more exhaustive kernel replay was still running as of Sunday, not yet finished.
Why mathematicians are not calling it solved yet
A machine accepting a proof is different from mathematicians understanding why it works. The Lean and Mathlib communities put it plainly in their joint statement: “We have full confidence that Alpoge’s proof of the Riemann Hypothesis is correct.” But that statement also says the new mathematical ideas behind the proof still need to be explained to, and checked by, the wider number theory community, a process that has not started.
The Clay Mathematics Institute’s own rules require a proof to appear in a peer-reviewed journal and then hold up for two years of open scrutiny before any Millennium Prize is considered, a bar the 1994 proof of Fermat’s Last Theorem cleared only after months of line-by-line human refereeing. The advocacy group Association for Human Mathematics has said it is gravely concerned about what an AI-assisted, code-only proof of this scale represents for the field.
Claude’s role in the proof
Alpoge has said he set the overall architecture of the argument while Claude worked through algebraic manipulations, trace formulas and technical lemmas, particularly in the proof’s final stretch. This builds on an earlier, narrower result: an unreleased Claude research build raised a known lower bound on the share of Riemann zeta function zeros proven to sit on the critical line from 41.6 percent to 67.2 percent, a result Anthropic mathematicians reviewed internally in August alongside outside reviewers before disclosing it.
Anthropic itself has not issued a company statement on the Riemann Hypothesis claim. The release comes in the same week Anthropic began trading on Nasdaq, and about a month after mathematician Terence Tao said publicly that he expected it would take until around 2050 before AI could meaningfully contribute to a problem of this difficulty.
What happens next
Alpoge has said he plans lectures, workshops and a human-readable synthesis paper to walk mathematicians through the argument, none of which have been published yet. Sutherland’s verification page notes additional, slower checks, including a full single-threaded kernel replay of the entire export, were still in progress as of Sunday afternoon, more than 40 hours after they began.
Until the broader number theory community has reviewed and accepted the underlying mathematics, the Riemann Hypothesis remains, formally, an open problem on the Clay Institute’s list.
Frequently asked questions
Does this mean the Riemann Hypothesis is officially solved?
Not yet. The Lean and Mathlib communities confirmed the code compiles and machine-checks against the stated theorem. The Clay Mathematics Institute requires peer-reviewed publication and two years of community acceptance before it considers a proof settled.
Does Levent Alpoge get the $1 million Millennium Prize?
Not at this stage. No prize has been awarded or announced, and the Clay Institute’s rules require the multi-year review process to run first.
Did Claude write the proof on its own?
No. Alpoge has said he directed the overall mathematical architecture, with Claude working through specific algebraic steps and technical lemmas under his direction.



Leave a Reply