Anthropic itself does not expect Claude's techniques to prove the Riemann Hypothesis. The conjecture that every nontrivial zero of the Riemann zeta function lies on the critical line, where the real part is one half, remains unproved. Anthropic's narrower claim is still substantial. An unreleased research version of Claude produced an unconditional proof raising the known lower bound for the share of zeros on that line from 41.6% to 67.2%. This is a bound on an asymptotic proportion, not a proof that all zeros lie there and not a statement about where the remainder lies.
The scale is the news. Anthropic's paper says the 41.6% record had stood since 2020. On the rounded figures in the announcement, Claude moved it by 25.6 percentage points in one result. Nothing in the historical progression laid out in the paper is close to that single jump. The proof draws heavily on pair-correlation work by Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh, along with Bombieri's earlier work. Its new move is a linear-algebraic reading of that machinery, treating zeros on and off the line together through a quadratic form. The result extends human mathematics; it did not appear without it.
The verification is stronger than a plausible model transcript. Anthropic released a Lean 4 formalization against Mathlib's definition of the zeta function, and says it passes the standard comparator with no unfinished proof placeholders. Anthropic mathematicians Levent Alpöge and Ralph Furman studied and validated the result. Brian Conrey and Dan Goldston examined the paper externally on short notice. Conrey's own work appears in the classic history of critical-line bounds, which makes his scrutiny meaningful. It is still important to name the limit: Anthropic released the model, the paper and the validation account. External examination is not the same thing as an independent peer-reviewed publication.
The human spark was unusually small and specific. Jarred Sumner, the creator of Bun and a non-mathematician, asked Claude to take a real stab at the hypothesis. Anthropic says the model first tried 650 ideas without success, then spent a day and a half coordinating about 60 subagents. Across 31 million output tokens, they ran 2,400 shell commands, checked candidate arguments and reviewed one another's work. Sumner's input was mostly messages of encouragement. Claude later searched for counterexamples, downloaded 54 papers to check novelty and independently reproved the finding. Those counts describe one expensive research run, not a repeatable benchmark.
The consequential part is neither the motivational anecdote nor the model brand. It is a concrete theorem that moves a century-old bound, arrives with a machine-checkable artifact and carries an explicit limit from its maker. Because the research model is unreleased, outside researchers cannot yet test whether the capability generalizes. Because the disclosure is a single-source Anthropic account, the broader mathematics community still has to work through the proof and its novelty. If the result survives that process, it belongs in the number theory record. It does not make the Riemann Hypothesis any less unsolved.