An unreleased research version of Anthropic's Claude model has achieved a notable advance in analytic number theory, improving a lower bound related to the Riemann hypothesis. The model increased the proven proportion of zeros of the Riemann zeta function that lie on the critical line from 41.6% to 67.2%. While this does not constitute a proof of the Riemann hypothesis itself, the gain represents the largest single-step advance on this specific bound in history.
The Riemann hypothesis, first proposed in 1859, is one of mathematics' most significant unsolved problems. It posits that all non-trivial zeros of the Riemann zeta function lie on a specific vertical line known as the critical line. This conjecture has profound implications for the distribution of prime numbers. Proving the hypothesis would confirm many existing mathematical results that rely on its truth.
Mathematicians have historically approached the problem by trying to prove that a minimum proportion of these zeros lie on the critical line. Before Anthropic's announcement, this lower bound stood at 41.6%, a figure achieved through decades of incremental work by mathematicians. The advance by Claude builds upon prior research, combining work from mathematicians such as Baluyot, Goldston, Suriajaya, Turnage-Butterbaugh, and Bombieri. Specifically, it enabled techniques introduced by Montgomery in 1973, which had previously assumed the hypothesis was true, to be applied unconditionally.
Anthropic emphasized that the model's techniques are not expected to lead to a full proof of the Riemann hypothesis. The gap between proving 67.2% and proving 100% is described as a different problem, not simply a matter of extending current methods. However, the process by which Claude arrived at the result is considered by some to be as significant as the numerical outcome. The model was tasked with an open-ended research problem, exploring hundreds of potential directions and coordinating multiple agents. It reportedly searched existing literature, performed numerical experiments, and reviewed its own work.
Two mathematicians at Anthropic studied and validated the generated proof, producing a concise note for experts. Additionally, Claude produced a formally verifiable proof in the Lean theorem prover, which passed standard validation tools. Experts Brian Conrey and Dan Goldston also reviewed the paper on short notice.
The development occurred over a 36-hour period, involving approximately 60 subagents and generating 31 million output tokens. This rapid progress contrasts with the typical pace of advancements in analytic number theory, which often occur in single percentage point increments. Anthropic has stated that the operator's input during the process was largely limited to messages of encouragement, which seemed to help the model overcome initial skepticism.
While the result has generated considerable attention, some reports have noted a lack of full corroboration from public primary sources or independent reporting, emphasizing the need for verifiable AI mathematical research. Anthropic has not released the specific model checkpoint or detailed process transcripts, which limits independent reproducibility of the AI's capability claim. However, the formalized proof is publicly available.