Article: Claude AI agents have lifted the verified proportion of Riemann-zeta zeros that lie on the critical line from 41.6 % to 67.2 %. The jump does not prove the Riemann hypothesis, but it shows that large-scale AI-driven workflows can produce a mathematically checkable result.
Why the Riemann hypothesis matters
The hypothesis asserts that every non-trivial zero of the Riemann zeta function has real part exactly ½. Those zeros control the distribution of prime numbers, so confirming the conjecture would tighten the link between prime patterns and complex analysis. To date, only a fraction of the zeros have been proved to sit on the “critical line.” Raising that fraction is a standard benchmark for progress.
How Anthropic got to 67.2 %
Anthropic ran Claude Code, its code-focused model, through two intensive research sessions. The second session deployed 60 sub-agents that coordinated via more than 2,400 shell commands and generated hundreds of Python scripts. In total the process emitted roughly 31 million output tokens—about the size of a short novel.
Mathematicians on the team audited the resulting paper and released a Lean 4 formalization, a machine-checked proof language, so others can verify each step. The work used an internal, unreleased research version of Claude, and verification relied on internal checks and brief expert reviews rather than formal peer review.
What the result actually says
The new paper raises the proven lower bound of zeros on the critical line from 41.6 % to 67.2 %. The figure is a lower bound: many more zeros may lie on the line but the analysis did not capture them.
Crucially, the result does not move the hypothesis itself any closer to a full proof. The hypothesis requires all non-trivial zeros to be on the line; a 67.2 % lower bound leaves more than a third of the zeros unaccounted for.
Limits of the AI approach
Anthropic’s own statements make clear they do not expect the current workflow to settle the hypothesis soon. Humans still have to review the model’s output for logical gaps, and the internal verification process lacks the rigor of community peer review. The research version of Claude is not publicly available, so independent replication is limited.
What comes next
The release of the Lean 4 formalization invites the broader mathematical community to audit the proof.
Takeaway
Anthropic’s Claude agents have demonstrated that AI can accelerate the verification of complex mathematical statements, delivering a sizable improvement in the known lower bound for Riemann-zeta zeros. The breakthrough is a proof-of-concept for large research-agent workflows, not a solution to the millennium-prize problem itself.