The result, stated plainly
The Riemann hypothesis says the nontrivial zeros of the zeta function sit on the critical line. Since nobody can prove that, mathematicians settle for bounds: guarantees that at least some percentage of zeros are where they should be. Anthropic's unreleased research model moved that guaranteed floor from 41.6% to 67.2%. On a number that usually moves in small increments when humans publish advances, 25 percentage points is a genuine jump.
Check the working
This is the part that separates the announcement from AI marketing. The result was formalized in Lean, the proof-assistant language that forces every logical step to be machine-checked. A natural-language proof sketch can hide subtle gaps. A Lean-checked proof compiles or it does not. Anthropic says the model explored around 650 distinct approaches, ran them through 60 coordinated subagents, and stitched the surviving thread into a verified result. Two in-house mathematicians reviewed it; outside experts were brought in on short notice before publication. That process costs real effort. It is also the same verification discipline behind Anthropic's Fermat's Last Theorem formalization from September, which reportedly ran to about 13 million lines of Lean and 29,500 intermediate theorems in under two weeks.
What the post doesn't claim
Anthropic has been unusually candid here, and the candor matters. The company says Claude didn't succeed at the challenge it was given. It did not prove the Riemann hypothesis. It does not collect the Clay Institute's $1 million bounty. And Anthropic positions the work as extending decades of existing mathematical research rather than conjuring a breakthrough from nothing. In a week where math announcements are competing for superlatives, that restraint is worth more than the number.
The other labs' math weeks
The contrast is the story. On October 6, OpenAI published 722 papers covering 377 open mathematical problems, produced by an unnamed internal frontier model and dumped on GitHub. Mathematicians are now combing through the batch for gaps, duplication, and problems that were already solved elsewhere. Anthropic published one tightly scoped number with a full methodology and a machine-checked proof. OpenAI bet on volume and left verification to the crowd. DeepMind took a third route years earlier: AlphaProof and AlphaGeometry 2 as specialized systems that reached silver-medal level at the 2024 International Mathematical Olympiad, competition problems rather than open research. Three labs, three experiments in how AI does mathematics. The interesting question is which approach the mathematical community ends up accepting as real work; right now, the approach that compiles is the one the mathematical community can actually build on.
The practical read
Math has become the new proving ground for frontier models, replacing coding leaderboards as the venue where labs show they can do something professionals recognize. For anyone watching that race, the metric to track is not parameter counts or paper counts. It is whether the claimed result survives machine checking. The Lean step is doing more work here than the headline: it is the difference between an impressive-looking output and a contribution the field can build on. Anthropic seems to have understood that earlier than its rivals.
Sources
- [1] Tech Insider report, "Anthropic Claude Hits 67.2% Riemann Zeta Bound [2026]" (October 7, 2026)Read source