Unreleased Anthropic Model Advances Progress on Riemann Hypothesis

Anthropic announced Monday that an unreleased internal model made significant progress on the Riemann hypothesis, one of mathematics' oldest unsolved problems, by significantly raising the lower bound of solutions for which the hypothesis is confirmed to hold true.

The Riemann hypothesis, first posed more than 150 years ago, concerns the distribution of prime numbers and carries a $1 million Millennium Prize for a full general proof — a prize that remains unclaimed. Anthropic's model did not solve the hypothesis outright, but according to the company, it meaningfully extended the range of values for which the hypothesis is known to be true.

How the result was produced

According to Anthropic, an employee without significant mathematical training prompted the model to "take a real stab" at the problem, then let it run largely unsupervised for roughly a day and a half. Over that period, the model:

• Tested 650 distinct approaches to the problem
• Coordinated work across 60 sub-agents
• Spent 31 million tokens in total (per Anthropic's disclosure)

A footnote in Anthropic's paper breaks down the sub-agents' roles: two agents developed the key mathematical ideas, 13 contributed supporting ideas, 30 attempted but failed to generate new ideas, 13 acted as validators checking the correctness of arguments, and two helped write the final paper.

Anthropic says the result was independently confirmed by two in-house mathematicians and formalized using Lean, an open-source proof assistant used to verify mathematical arguments line by line. None of these findings have been independently peer-reviewed by the broader mathematical community as of publication.

Part of a broader pattern

This result follows a string of AI-assisted mathematical claims in 2025 and 2026. Multiple Erdos problems have reportedly been resolved with AI assistance this year, and OpenAI recently claimed ten major results from an internal model referred to as "Astra." A separate Anthropic effort earlier disproved the long-standing Jacobian conjecture.

The growing volume of AI-assisted results has split the mathematics community. In June, a group of prominent mathematicians signed a public declaration warning that AI-generated proofs could erode a core norm of the field: that mathematical proofs be attributable to identifiable authors who take responsibility for their correctness.

Fields Medal winner Timothy Gowers pushed back on that framing in a blog post, suggesting the shift might not be entirely negative. "If we arrive at a world where mathematical theorems are no longer associated with mathematicians, maybe that won't be any more problematic than the fact that stars aren't named after astronomers and most aren't named at all," Gowers wrote.

What this means

Anthropic has not disclosed which model produced this result, its parameter count, or when — or if — it will be released, so the claim cannot be independently verified against a public system. The Lean formalization and confirmation by in-house mathematicians lend some credibility, but "in-house" verification is not the same as peer review, and Anthropic has a commercial interest in showcasing its unreleased model's capabilities.

What's harder to dispute is the trend: AI systems are increasingly being used as active collaborators in frontier mathematical research, not just as calculators or proof-checkers. Whether that's cause for celebration or a structural challenge to how mathematics assigns credit and responsibility remains an open question — one the field itself is visibly struggling to answer.