Anthropic reports Claude progress on a Riemann zeta bound
AI math just got a number you can argue with. On August 10, Anthropic said an unreleased research version of Claude raised a longstanding lower bound on the share of Riemann zeta zeros that lie on the critical line — from 41.6% to 67.2%. That is not a proof of the Riemann hypothesis. It is a concrete advance on a related constant mathematicians have been inching upward for decades.
What bound moved, and by how much
The Riemann hypothesis says the non-trivial zeros of the zeta function all sit on a specific vertical line. Nobody has proved or disproved it since 1859. Parallel work asks a weaker question: what minimum fraction of zeros can we prove sit on that line?
Anthropic says Claude improved that lower-bound proportion from 41.6% to 67.2%, drawing on prior analytic number theory. The key combination joins results from Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh with a 2000 paper by Bombieri, in a line that lets Montgomery-style techniques operate without assuming the hypothesis. Claude’s technical move, in Anthropic’s short sketch, treats a Weil-induced quadratic form over a suitable function space, accounts for positive- and negative-definite subspaces from zeros on and off the line, then writes a rank inequality from first- and second-moment information.
External experts Brian Conrey and Dan Goldston examined the paper on short notice. Anthropic mathematicians Levent Alpöge and Ralph Furman validated the write-up and produced an informal note for experts. Vendor-hosted validation is not journal peer review — but the bound is a falsifiable mathematical claim, not a vibe benchmark.
How Lean formalization figured in
Beside the informal paper, Claude worked with Anthropic staffer Eric Easley to produce a Lean formalization. Anthropic says it passes the standard validation tool comparator. That matters more than blog drama: a machine-checkable artifact lets skeptics inspect the proof object instead of arguing about chat transcripts.
The discovery path was agentic and expensive. Anthropic says the unreleased Claude found the bound across two Claude Code sessions totaling about 31 million output tokens. Staffer Jarred Sumner (not a mathematician) prompted the model to “take a real stab” at the hypothesis. After 650 failed ideas, a second pass coordinated roughly 60 subagents, ran about 2,400 shell commands, wrote hundreds of Python scripts, checked numerical zeros, downloaded 54 arXiv papers looking for prior art, and re-proved the finding from scratch.
What the model failed to prove
Claude did not prove the Riemann hypothesis. Anthropic states that up front and says it does not expect Claude’s techniques to yield a full proof. The million-dollar Clay problem remains open. The model was pointed at the hard target and, while failing there, improved a related lower bound that human mathematicians had already framed.
That distinction is the whole story. A bound jump from 41.6% to 67.2% is research progress. A solved RH would be a century-scale event. Conflating the two is how AI math coverage turns into mythology. Honest frame: Claude extended existing analytic machinery far enough to move a published constant, then helped package the argument for Lean checking.
Why mathematicians will argue authorship
Authorship fights are inevitable when a non-mathematician prompts a model, subagents generate the lemmas, lab mathematicians validate, outside experts glance, and Lean absorbs the formalization. Who “found” the bound — Sumner’s prompt, Claude’s search, the prior papers Claude combined, or the humans who certified the argument?
Analysis: credit should track the mathematical content. The novelty claim is the combination and the inequality that lifts the constant; prior literature supplied the tools. Lean lowers the trust tax but does not settle credit. Watch whether independent number theorists reproduce the 67.2% bound from the paper and formalization — without anyone pretending RH fell in a chat window.
