Claude raises Riemann zero bound to 67.2% in 31M-token run

Anthropic's unreleased research version of Claude improved the proven lower bound for Riemann zeta function zeros on the critical line from 41.6% to 67.2%, a 25.6-point jump that mathematicians had advanced by only 0.8 points over the previous 37 years.

"Claude's result is extraordinary — it may be the most significant breakthrough in analytic number theory since the bounded prime gap result in 2013," said Deedy, a partner and researcher at venture capital firm Menlo Ventures.

The result emerged from two Claude Code sessions consuming 31 million output tokens. Claude generated and tested 650 failed approaches before coordinating roughly 60 sub-agents that executed 2,400 shell commands, wrote hundreds of Python scripts, and ran thousands of numerical checks against known zeta zeros. Anthropic mathematicians Levent Alpöge and Ralph Furman validated the paper, while number theorists Brian Conrey and Dan Goldston reviewed it. Claude also produced a Lean formal proof verified through the standard comparator tool.

The advance demonstrates frontier AI models can engage with genuinely open research problems. While Anthropic said the techniques won't directly prove the Riemann hypothesis — a 167-year-old conjecture carrying a $1 million Clay Mathematics Institute prize — the result signals accelerating AI mathematical capability that could reshape how research institutions allocate compute budgets and how AI labs position their models for scientific discovery.

A 167-Year-Old Conjecture Meets 60 Sub-Agents

The Riemann hypothesis, posed by Bernhard Riemann in 1859, asserts that all non-trivial zeros of the zeta function lie on the critical line where the real part equals 1/2. The conjecture underpins dozens of results about prime distribution, and its proof or disproof carries a $1 million Millennium Prize from the Clay Mathematics Institute.

Claude's approach built on work by Hugh Montgomery from 1973, which introduced techniques for studying zero distribution but assumed the hypothesis was true. Recent papers by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh removed that assumption, and Claude combined those results with a 2000 paper by Enrico Bombieri.

Technically, Claude constructed a function space with a Weil-induced quadratic form, mapping zeros on and off the critical line to positive- and negative-definite subspaces. The key step was treating the entire space together rather than separating the positive and negative parts, while allowing the quadratic form to be non-diagonal. This unified treatment, combined with first- and second-moment information, produced the inequality that yielded the 67.2% bound.

How Claude Worked Through 650 Failed Ideas

Jarred Sumner, an Anthropic staff member with no mathematics background, prompted Claude to "take a real stab" at the hypothesis. After 650 failed ideas, Sumner's encouragement — mostly "keep going" and "believe in yourself" — helped Claude push through initial skepticism. The model then self-verified by having sub-agents review proofs, search for counterexamples, download 54 arXiv papers to check for prior results, and independently re-derive the finding.

Anthropic emphasized that the approach is not expected to lead directly to a final proof of the Riemann hypothesis. The 67.2% figure represents an improved lower bound on a related problem, and there remains a vast theoretical gap between this result and a full proof.

Investor Implications for AI Research Spending

The result shows AI models can extend the reach of human mathematical ideas in unexpected ways, potentially accelerating discovery across fields that depend on number theory, including cryptography and computational complexity. For investors, the demonstration strengthens the case for continued AI R&D spending at frontier labs, even as questions persist about when such capabilities translate into commercial products.

Anthropic, valued at over $180 billion in its latest funding round, competes with OpenAI and Google DeepMind in the race to push model capabilities beyond current benchmarks. The Riemann result adds to a growing list of AI-assisted scientific discoveries, from cryptographic weakness identification to drone control, that could influence how enterprises and governments evaluate frontier model procurement.

This article is for informational purposes only and does not constitute investment advice.