Anthropic's unreleased AI attacked the Riemann hypothesis for 36 hours and found a new lower bound

For 150 years, the Riemann hypothesis has been the Everest of mathematics—so famously unsolved that the Clay Mathematics Institute hung a $1 million bounty on it. Anthropic just lobbed an AI at the summit. An unreleased model, guided by a researcher with no advanced math background, chewed on the problem for a day and a half and pushed the proven lower bound significantly higher. That’s the kind of incremental, granular progress that real working mathematicians make, not parlor tricks from a chatbot.

The process was less a stroke of genius and more a relentless campaign. The model orchestrated 60 sub-agents to test 650 different lines of attack, burning through 31 million output tokens. Two of those sub-agents cracked the key mathematical ideas, while a swarm of others validated the arguments. Two in-house mathematicians at Anthropic confirmed the results, and the whole thing was formalized using the open-source proof assistant Lean. It’s a deeply weird research paradigm—a swarm intelligence producing verifiable truth without a single human ever truly understanding the full arc of the creative process.

This isn’t happening in a vacuum. AI models have been gnawing at the edges of pure math all year, solving Erdos problems and, in OpenAI’s case, generating ten major results from their internal “Astra” model. Anthropic itself recently used AI to disprove the Jacobian conjecture. The Riemann result, however, lands with a different kind of weight because of the problem’s mythic status. It forces a conversation the field has been dodging.

That conversation is already fractious. A group of prominent mathematicians signed a declaration this June arguing that proofs must be attributable to humans who can claim credit and shoulder blame. It’s a defense of authorship as a moral act. But Fields Medalist Timothy Gowers offered a characteristically lucid counterpoint: if theorems eventually lose their association with specific people, he wrote, it might be no more tragic than the fact that we don’t name stars after astronomers. The real anxiety isn’t about truth—it’s about the obsolescence of the truth-teller. And after this week, that anxiety has a new lower bound, too.