Unreleased Claude research model improves a Riemann hypothesis lower bound from 41.6% to 67.2%

An unreleased research version of Claude has extended a long-standing partial result tied to the Riemann hypothesis, one of mathematics' most famous unsolved problems, according to a paper Anthropic published August 10, 2026. Working across two sessions totaling roughly 31 million output tokens, the model raised the proportion of zeros of the Riemann zeta function proven to satisfy the hypothesis from 41.6% to 67.2%.

What's new

Anthropic's post states plainly: "An unreleased research version of Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis."

The model did not prove the Riemann hypothesis itself — a $1 million Clay Millennium Prize problem that has resisted proof for over 160 years. It first attempted a direct proof and failed. But in the process it found a way to combine several existing partial results into a stronger bound. As Anthropic describes it: "Claude found that combining the results from Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh with the work of Bombieri provides a way to surpass the previous state-of-the-art lower bound proportion of 41.6%, increasing it to 67.2%."

The scale of the search behind that result is notable. The model coordinated roughly 60 subagents that together ran about 2,400 shell commands and hundreds of Python scripts, and generated 650 unsuccessful lines of attack before landing on the winning combination of existing theorems.

Context

The Riemann hypothesis concerns where the "non-trivial" zeros of the Riemann zeta function fall — a question tightly linked to how prime numbers are distributed. Since a full proof has eluded mathematicians for over a century, one active research program instead tries to prove that some large fraction of the zeros satisfy the hypothesis, tightening the bound over time through incremental theorems. The 41.6% figure Claude surpassed represented the prior state of the art, built on work by mathematicians including Baluyot, Goldston, Suriajaya, Turnage-Butterbaugh, and Bombieri. Claude's contribution wasn't a new theorem from scratch — it was recognizing a productive way to combine those existing pieces that human researchers hadn't assembled.

This lands alongside a string of other recent examples of frontier models being pointed at open problems in pure mathematics rather than benchmark suites, as labs test whether large-scale automated search and reasoning can generate genuinely new research contributions rather than just replicate known results.

Why it matters

A computer-assisted advance on a named constant tied to the Riemann hypothesis is a concrete, checkable data point in the debate over whether frontier models can do original mathematical research rather than pattern-match existing proofs. Anthropic's own framing is measured but pointed: "This result shows that AI models like Claude can extend the impact and reach of mathematicians' ideas in new and sometimes surprising ways."

The process also illustrates what this kind of AI-driven research currently looks like in practice: not a single flash of insight, but brute-force parallel search — dozens of subagents, thousands of shell commands, hundreds of dead ends — that eventually surfaces a combination a human might have found faster with the right intuition, or might never have tried at all. That raises open questions for the mathematics community about how to verify, credit, and build on results produced this way, especially as more labs point unreleased frontier models at other unsolved problems.