AI Just Nudged the 150-Year-Old Riemann Hypothesis Forward

Sixty AI subagents, 31 million output tokens, about a day and a half — that’s the entire toolkit an as-yet-unnamed, unreleased Anthropic model used against the Riemann hypothesis, a problem that’s stumped mathematicians for more than 150 years and still carries an unclaimed $1 million bounty from the Clay Mathematics Institute. The model didn’t solve it. What it did was significantly push up the lower bound of solutions for which the hypothesis holds true — and it’s the third time in 2026 alone that AI has made progress on pure mathematics serious enough for mathematicians to actually take note of.

What the nameless model actually did

Anthropic laid out the division of labor in a footnote to its paper: of the 60 subagents, two were responsible for developing the key mathematical ideas, 13 contributed ideas to the lead agent, 13 verified that the arguments held up, and two helped write the paper — the remaining 30 tried and failed to come up with anything new. Across the whole run, the model tested 650 different ideas, produced 31 million tokens of output, and took roughly a day and a half.

This isn’t just Anthropic’s word for it. The result was confirmed by two in-house Anthropic mathematicians and formalized using Lean, the open-source proof assistant mathematicians use to verify that a piece of reasoning is actually airtight — not something Anthropic gets to self-certify. It’s worth being precise about what was actually achieved: the model significantly raised the lower bound of solutions consistent with the hypothesis, not the hypothesis itself. The Clay Institute’s million-dollar prize is still unclaimed.

2026 is turning out to be a big year for AI and pure math

This isn’t an isolated case. OpenAI’s internal build of its “Astra” model recently solved ten previously open problems spanning mathematics and theoretical computer science. Anthropic separately disproved the long-standing Jacobian conjecture. Several well-known Erdős problems have also been resolved by different AI models this year. 2026 is shaping up to be the year AI started showing up seriously in pure math research, and the Riemann hypothesis is just the loudest target so far.

What mathematicians make of it: cautious optimism and an open letter

The reaction hasn’t been uniform. Fields Medalist Timothy Gowers struck a relatively optimistic note, arguing that if mathematical theorems stop being tied to a specific mathematician’s name, that might not be any more of a problem than the fact that stars aren’t named after astronomers — and most stars don’t have names at all.

But in June, a group of mathematicians published an open letter (see leidendeclaration.ai) raising concerns that AI could undermine some of the field’s core values — particularly the long-assumed standard that a genuine mathematical proof should be attributable to a specific author. Anthropic’s model, which doesn’t even have a name, sits right in the middle of that debate: if no one is willing — or able — to put their name behind an AI model’s mathematical contribution, whose result is this push on the Riemann hypothesis’s lower bound, exactly?

The Riemann hypothesis is still the Riemann hypothesis, and the Clay Institute’s million dollars is still sitting unclaimed. But after watching AI rack up one verifiable result after another on pure math’s hardest open problems this year, “mathematical proof can only come from a human hand” isn’t a claim anyone seems eager to make with much confidence anymore.