An unreleased Anthropic Claude model achieved a mathematical breakthrough on the Riemann hypothesis, raising the known lower bound for Riemann zeta function zeros from 41.6 percent to 67.2 percent after an internal researcher encouraged the AI to continue its exploratory work.
Artificial intelligence models are moving beyond answering questions and actively participating in the research process itself. In an internal experiment conducted by Anthropic, an unreleased version of the Claude AI model stumbled into a notable mathematical result while attempting to tackle one of mathematics’ most famous unsolved problems.
The target of the experiment was the Riemann hypothesis, a problem introduced in 1859 that remains one of the Clay Mathematics Institute’s Millennium Prize Problems. The institute carries a $1 million reward for a verified proof of the hypothesis. While negative even numbers are known to yield a zero output in the Riemann zeta function, the hypothesis posits that all other zeros exist without exception along a specific straight line.
How Claude Approached the Riemann Hypothesis
The experimental process began when an Anthropic staff member with no significant mathematical training prompted the model to take a real stab
at the hypothesis, according to company disclosures. Operating within a self-contained AI agent service, the research model generated 31 million output tokens across two Claude Code sessions, working autonomously for approximately a day and a half.
The initial phase proved difficult. Claude generated 650 distinct ideas, all of which failed to produce a viable path forward. Encountering this initial wall, an Anthropic team member intervened with a simple message telling the AI not to give up and urging it to believe in itself. Encouraged by this human prompt, the system coordinated approximately 60 subagents to run thousands of numerical checks, execute 2,400 shell commands, and write hundreds of Python scripts.
Raising the Lower Bound on Zeta Function Zeros
Claude did not solve the Riemann hypothesis, and Anthropic officials stated plainly that they do not expect the techniques used by the model to lead to a full proof. Instead, the model achieved a byproduct discovery regarding the proportion of zeros lying along the designated straight line.
Previous mathematical research established that at least 41.6 percent of these zeros lie on the line. By combining several existing research frameworks developed by mathematicians including Baluyot, Goldston, Suriajaya, Turnage-Butterbaugh, and Bombieri, Claude discovered a method to raise that lower bound to 67.2 percent.
The model also produced a formally verifiable proof using the Lean proof assistant, adding a layer of computational verification to the discovery.
Expert Review and Verification Process
Because mathematical claims involving AI require rigorous oversight, Anthropic subjected the findings to internal and external scrutiny. The results were reviewed by company mathematicians Levent Alpöge and Ralph Furman, and the company shared the paper with outside experts Brian Conrey and Dan Goldston.
Independent computer verification was also completed.
Broader Implications for Scientific Research
The experiment highlights how advanced language models can function as exploratory collaborators rather than mere query engines. Researchers across scientific and technical fields spend enormous amounts of time testing hypotheses, writing code, and eliminating dead ends.
While the achievement is narrow and has not yet cleared standard academic peer review, the experiment demonstrates that systems like Claude can help human researchers investigate a broader array of mathematical possibilities, signaling a shift in how computational tools assist theoretical research.
Worth a look
