The Riemann hypothesis has remained one of mathematics’ major unsolved problems for more than 150 years, concerning the distribution of prime numbers. A general, working proof carries a current prize of $1 million, which remains unclaimed. While present AI models do not offer a universal solution, Anthropic announced on Monday that an unreleased internal model made significant progress by increasing the lower bound of inputs for which the hypothesis has been verified.
According to Anthropic, the effort began when an internal staff member without significant mathematical training instructed the model to “take a real stab” at a proof. The staffer then left the system to operate, and over the following day and a half the model coordinated the work itself.
Anthropic reports that the model tested 650 different ideas, coordinated activity across 60 subagents, and produced roughly 31 million output tokens. A footnote in the company’s writeup breaks down the roles of those subagents: two developed the key mathematical ideas; 13 contributed ideas to those agents; 30 attempted but failed to develop new ideas; 13 acted as validators checking correctness of arguments; and two helped draft the initial paper.
The company says two of its in‑house mathematicians confirmed the finding, and the result was formalized using the open‑source proof assistant Lean. Anthropic has not specified when or whether the model will be released for broader use.
Context within AI‑driven mathematics
This announcement follows a string of results in which large language models (LLMs) have played a central role in mathematical advances. Over the past year, several Erdős problems were reported as solved with AI involvement, and more powerful models have produced increasingly notable outcomes. OpenAI recently published ten major results attributed to its internal “Astra” model, and a separate effort from Anthropic reportedly refuted the longstanding Jacobian conjecture.
These developments have generated both enthusiasm and concern within the mathematics community. In a public declaration signed in June, a group of prominent mathematicians warned that AI might undermine core values of the field—particularly the idea that correct mathematical proofs should be attributable to specific authors who accept credit and responsibility for their work.
At the same time, opinion is divided about how the discipline should adapt. In a blog post responding to that declaration, Fields Medal winner Timothy Gowers questioned whether AI’s influence might alter mathematics in more complex and potentially positive ways. Gowers suggested that detaching theorems from individual mathematicians may not be more problematic than other naming conventions in science, such as the fact that most stars are not named after astronomers.
Significance and next steps
Anthropic’s reported advance does not amount to a general proof of the Riemann hypothesis, but it does extend the verified range for which the hypothesis holds. The episode demonstrates that large language models can coordinate multi‑step mathematical research processes and generate substantive ideas even when initial prompting comes from a user without deep mathematical expertise. That capability intensifies ongoing debates about authorship, responsibility, and how AI should be integrated into mathematical practice.
Anthropic has yet to release fuller technical details, the formalized proofs themselves, or a timeline for public access to the model; further disclosures will be necessary for the community to independently evaluate and build on the work.
