The day after Anthropic’s record IPO, a mathematician at the company published a four-page note on his personal site. “I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis,” Levent Alpöge wrote. He linked the work in a footnote: 14 pieces of a compressed archive containing the argument in Lean, a language mathematicians use to write proofs a computer can check.

Bernhard Riemann proposed the hypothesis in 1859. It tells mathematicians how precisely they can predict the distribution of prime numbers, and the Clay Mathematics Institute has put $1 million behind a solution. Alpöge says his method also settles a broader version of the problem. As of Friday morning, we could find no independent confirmation of the full proof.

Alpöge says he had the result on Sunday

On Thursday, Anthropic priced 980 million shares at $83 apiece in the biggest IPO on record. Alpöge addressed the timing on the second page of his note. “The precise timing is horrible,” he wrote. By his account, he reached the result late Sunday and messaged Anthropic CEO Dario Amodei to ask when he could announce it. He says there was a constraint on the announcement without specifying what it was.

Alpöge had checked Claude’s work on this very problem before. In August, Anthropic reported that Claude had fallen short of proving the hypothesis but had produced a narrower result. Alpöge and fellow mathematician Ralph Furman examined that earlier work for the company. This time Alpöge is making the claim himself, and he credits Claude with doing mathematics near the end that he is still working to understand.

What outsiders have checked so far

By Friday morning, an independent verification project had matched its copy of the files to Alpöge’s release and checked that the statement it wanted to test compiled against Lean’s standard mathematics library. For the Riemann hypothesis itself, the reviewers could use a definition already in that library. They also spelled out the scope of the broader claim: it covers a family of functions known as Dirichlet L-functions.

The project still listed the full check of Alpöge’s submission as pending. That’s the run that tests whether the steps in those 14 files actually prove the statements the reviewers specified, using only the permitted assumptions. Alpöge says he intends to write an account his colleagues can follow. For now, the code is the version they have to judge.