Anthropic’s Secret AI Model Just Cracked Part of Math’s 150-Year-Old Riemann Hypothesis

For more than 150 years, the Riemann hypothesis has sat at the top of mathematics’ “most wanted” list — a deceptively simple-sounding puzzle about how prime numbers are distributed, with a $1 million bounty attached to whoever proves it. This week, an unreleased AI model from Anthropic didn’t solve it outright, but it made real, verified progress. And that alone is enough to send a jolt through the math world.

What Anthropic’s Model Actually Did

On Monday, Anthropic revealed that an as-yet-unreleased model had significantly increased the lower bound of solutions for which the Riemann hypothesis holds true. In plain English: it pushed the boundary of what we can confirm further than before. It’s not a full proof — the $1 million prize remains unclaimed — but it’s a genuine mathematical result, not a party trick.

The Most Surprising Part: How It Happened

Here’s the detail that has everyone talking. The breakthrough wasn’t driven by a team of PhD mathematicians. An Anthropic staff member without significant mathematical training simply prompted the model to “take a real stab” at the problem — then left it running for about a day and a half. The AI did the rest, coordinating the entire effort on its own.

650 Ideas, 60 Subagents, 31 Million Tokens

The scale of the effort is staggering. The model tested 650 different ideas, coordinated across 60 subagents, and burned through 31 million output tokens. According to Anthropic, only two of those subagents developed the key mathematical ideas, 13 fed them supporting ideas, 30 tried and failed, 13 acted as validators checking the logic, and two wrote up the initial paper. It’s a glimpse of AI working less like a chatbot and more like an autonomous research lab.

Not the First — and Not the Last

This isn’t a one-off. AI models have been quietly racking up mathematical wins all year. Several long-standing Erdős problems have fallen to AI in 2026, OpenAI recently published 10 major results proved by its internal “Astra” model, and a separate Anthropic effort disproved the decades-old Jacobian conjecture. Each new, more capable model seems to push the frontier a little further.

Why Mathematicians Are Both Thrilled and Worried

The reaction inside the math community is split. In a public declaration signed earlier this year, a group of prominent mathematicians warned that AI could erode a core value of the field — the idea that proofs should be attributable to specific authors who take credit and responsibility for their correctness. If a swarm of AI subagents produces the proof, who exactly gets the credit?

Not everyone sees a crisis. Fields Medal winner Timothy Gowers offered a more open-minded take, suggesting that a world where theorems aren’t tied to individual mathematicians might not be a disaster — comparing it to how most stars aren’t named after the astronomers who found them. The field, clearly, is still figuring out where it stands.

The Bottom Line

Anthropic’s model didn’t win the $1 million prize, and the Riemann hypothesis is still officially unsolved. But the takeaway is bigger than any single result: an AI, pointed at one of the hardest problems humans have ever posed and left largely to its own devices, made verifiable progress that in-house mathematicians confirmed and formalized in Lean. Whether that excites you or unsettles you, one thing is clear — AI is no longer just doing our math homework. It’s starting to do the kind of research we thought only humans could.

About the Author

JACK POTE

Administrator

Visit Website View All Posts