Anthropic published a result on 10 August in which an unreleased research version of Claude raised a lower bound in analytic number theory: the proportion of non-trivial zeros of the Riemann zeta function proven to lie on the critical line moved from 41.6% to 67.2%.
The subagent accounting
The work ran as a swarm of 60 subagents inside Claude Code. Anthropic breaks down what each did: 2 produced the key ideas, 13 contributed supporting ones, 30 pursued approaches that went nowhere, 13 validated arguments and 2 helped write the paper. It burned 31 million output tokens and 2,400 shell commands in about a day and a half.
Who checked it
Anthropic's own mathematicians Levent Alpöge and Ralph Furman validated the argument, with external review from number theorists Brian Conrey and Dan Goldston. That external step is what separates this from the usual claim of machine-generated mathematics.
What it is not
It is not progress toward proving the Riemann hypothesis, and Anthropic says so directly: it does not expect these techniques to lead to a proof. The bound improvement was a byproduct of attempting a different problem. Second-hand coverage has also added figures that are not in Anthropic's write-up, including "650 solution approaches" and "36 hours", and has aged the problem at 150 or 160 years — Riemann's paper is from 1859.
The interesting part is the ledger
Thirty of sixty agents failed, and that ratio is published. It is the first time a lab has shown the internal hit rate of a research swarm rather than only its output.
Nobody can rerun it
The system that produced this is unreleased, and Anthropic has not said whether it will ship. The result is unreproducible outside the company: the external reviewers checked the mathematics, not the claim about how it was generated.