Anthropic mathematician reports Claude-assisted proof of 1859 Riemann hypothesis

On November 20, 2026, Levent Alpöge, a mathematician at Anthropic, posted a note saying he believes he and Claude found a proof of the Riemann hypothesis, a conjecture Bernhard Riemann stated in 1859. The note says the method also reaches the Generalized Riemann Hypothesis, a wider family of functions, and Alpöge hopes to reach the Extended Riemann Hypothesis with more time. According to The Decoder, the formal proof runs to more than 250,000 pages. Alpöge published the argument only as Lean 4 code, which a computer can check line by line. CASRAI describes a 14-part archive of 175,187 Lean files that expands to more than 61.5 gigabytes, so the page and file counts measure different things.
Hours later, Ravi Vakil, a Stanford professor and president of the American Mathematical Society, posted a draft proof of the Generalized Riemann Hypothesis by Stanford postdoc Alessandro Masullo, calling it all human with no AI. Vakil had planned a year of polishing before going public, but he released the current draft after Alpöge's announcement. About eight hours later, Vakil updated his post to say several people had reported problems, some found with AI models, and that all must be resolved before the draft could be accepted.
On November 21, the Lean and Mathlib communities said they had full confidence that the proof is correct, after confirming its theorem matches Mathlib's statement of the Riemann hypothesis. MIT's Andrew Sutherland separately compiled the Lean files without problems. Human experts have not finished reviewing the argument, and Anthropic and Stanford had not commented in early reports.
Why it matters
For builders, the practical lesson is that the claim is being judged by its artifact rather than by the author's employer. Anyone can download the Lean files, compile them and check whether the stated theorem matches the established statement, which is what the Lean and Mathlib communities did within a day. That shifts diligence from reputation to reproducibility. Teams shipping AI-generated research, code or analysis should expect the same test: outputs a third party can rerun, with prompts, model versions and verification steps recorded.