If this holds up, it is the biggest result an AI system has ever helped produce, and it didn’t come from a lab demo. Levent Alpöge, a mathematician who works at Anthropic, claims he used Claude to prove the Riemann Hypothesis.
According to the BBC, as relayed by Engadget, Alpöge spent six weeks working with Claude to build a detailed proof. He called the problem “the sacred object in mathematics for me since my childhood.” Bernhard Riemann proposed it in 1859. The idea, simply put, is that the locations of prime numbers along the number line follow a meaningful order. Scientific American calls it often the most important unsolved problem in math, and the Clay Mathematics Institute lists it among its Millennium Prize Problems, with $1 million offered for the first correct solution.
The format matters more than the headline. Alpöge presented the work in Lean, a programming language used to write and verify mathematical proofs. That is the key detail. A Lean proof can be checked by machine, so the question is less “do we trust the AI?” and more “does the code compile, and did they formalize the right statement?” He says the full proof runs past 250,000 pages. No human referee is reading that cold, which is exactly why formal verification is the only realistic route.
James Maynard, an Oxford mathematician, told The New York Times he is “98 or 99 percent certain” the proof is accurate. That’s a serious endorsement from someone who knows the field. But it isn’t confirmation, and the Clay Institute has its own process for recognizing solutions.
The context is a crowded race. In October, OpenAI posted hundreds of solutions to previously unsolved math problems, plus partial progress on others, the Riemann Hypothesis included. Google DeepMind has also pushed hard on formal math with its AlphaProof work. Math has become the favored proving ground for reasoning models because answers can be checked, unlike most benchmarks, which are easy to game.
Two cautions. First, Alpöge works at Anthropic, so the company has an obvious interest in the story. Second, OpenAI’s batch of claims still needs verification too, so nobody has a clean win yet.
Does it matter? Yes, if the Lean check passes. A verified proof would show that a frontier model can be a real research partner on problems that beat the best human minds for 167 years. For developers and founders, that is a stronger signal than any leaderboard score. If it fails, it’s a useful reminder that big claims need big audits.



