A staff member at Anthropic with no serious math background typed a prompt asking an internal model to ‘take a real stab’ at the Riemann hypothesis, then walked away. A day and a half later, the model had made measurable progress on one of the most notoriously difficult problems in the history of mathematics. According to TechCrunch, Anthropic confirmed the result using two in-house mathematicians and formalized it with Lean, the open-source proof assistant used to verify mathematical arguments with machine precision.
The Riemann hypothesis has been open since 1859. It concerns the distribution of prime numbers and carries a $1 million Millennium Prize for a general proof, which nobody has claimed. The Anthropic model did not solve it. But it significantly pushed the lower bound of cases for which the hypothesis holds, which is not nothing. In mathematical terms, that’s real forward movement on a problem that has defeated every serious attempt for over a century.
What makes this worth paying attention to is the architecture behind the result. The model coordinated 60 sub-agents and tested 650 distinct approaches, spending 31 million tokens across the session. Of the 60 sub-agents, two developed the core mathematical ideas, 13 contributed supporting concepts, 30 failed to generate new ideas, 13 acted as validators checking argument correctness, and two helped draft the initial paper. That kind of internal scaffolding, running autonomously over 36 hours from a single casual prompt, is a meaningful signal about where agentic AI systems are heading.
This is not an isolated result. The math benchmark keeps moving. Several Erdos problems have fallen to AI models this year. OpenAI recently published ten results proved by its internal ‘Astra’ model. A separate Anthropic effort already disproved the Jacobian conjecture, a long-standing open problem. The pace is accelerating as models get more capable, and the Riemann result suggests the ceiling is higher than most researchers assumed.
The mathematical community has noticed, and it’s not uniformly enthusiastic. A public declaration signed in June by a group of prominent mathematicians argued that AI-generated proofs threaten a core value of the field: that proofs should be attributable to specific people who take intellectual responsibility for them. It’s a serious concern, not a reflexive one.
But Fields Medal winner Timothy Gowers pushed back in a blog post, suggesting the question might be more complicated than it looks. He compared the situation to stars, most of which carry no human name at all, without anyone treating that as a crisis. The field is genuinely split, and the Riemann result will sharpen that debate rather than settle it.
For developers and founders building on top of frontier models, the practical takeaway is straightforward: multi-agent orchestration running over extended time horizons is now capable of producing outputs that hold up to expert scrutiny in hard scientific domains. That matters well beyond pure mathematics.
