Anthropic mathematician Levent Alpöge: proving the Riemann Hypothesis with Claude
Anthropic technical team mathematician Levent Alpöge posted a four-page note on his personal website alpo.ge on November 20, 2026 (remark.pdf), saying that, in collaboration with Claude, he had discovered a proof of the Riemann Hypothesis and extended it to the generalized Riemann Hypothesis (GRH) for Dirichlet L-functions. The proof has been formally completed in Lean without additional assumptions; the human-readable text exceeds 250,000 pages, so only the Lean code was released (in 14 compressed packages, about 175,187 files and 17.4 million lines).
Princeton AI researcher Yifan Zhang reposted the note on X that same day and said it had been solved; Tao Zhe-xuan's blog the same day listed this result as one of the day's AI-assisted proofs with the lowest level of digestion, pointing out that it is a huge Lean-only proof, with preliminary analysis involving the Langlands program and arithmetic geometry, and directing readers to the Lean Zulip discussion thread. MIT's Andrew Sutherland independently compiled it successfully and called it a historic moment; another independent verifier reported that checks passed using stock Lean and an independent Rust kernel, relying only on the standard axioms.
The work timeline ran from October 6 to November 16 (about 41 days), starting with the Stepanov method and a theta-lifting chain (Claude suggested a length of about √X). Alpöge said that Claude pulled far ahead in the closing stage, while he himself had limited mastery of higher automorphic forms; the announcement came shortly after Anthropic's IPO, and he had written to Dario Amodei and admitted the timing was bad, but denied deliberately coordinating it. Clay Mathematics Institute rules require at least two years of review after formal publication before a million-dollar prize could possibly be considered.
ABAB AI Insight
The formal proof shifts mathematical verification from human reading to machine-checkable code closure, moving the bottleneck from discovery to understanding and simplification. Anthropic released this result the day after its IPO, turning internal AI capability into a publicly compilable output and changing the coupling between mathematical priority and capital narratives — proof rights no longer depend on journals, but on reproducible Lean archives and independent kernel verification.