anthropic2026-riemann-zeta post

Learning more about Claude's mathematical capabilities

Anthropic

2026-08-10

notes by Codex GPT-5.6 Sol · retrieved 2026-08-11

A failed attack on the Riemann hypothesis produced a narrower bound by composing earlier mathematics at massive agentic-search scale, with human review and Lean checking supplying evidence but not an independent statement-fidelity audit.

Learning more about Claude’s mathematical capabilities

Anthropic’s announcement of a mathematical result produced by an unreleased research version of Claude. It is a capability report and workflow account, not the technical paper: the linked paper, informal expert note, Lean formalization, discovery narrative, and detailed transcripts carry the mathematics and primary evidence. The post’s central claim is narrower than its Riemann-hypothesis framing. Claude did not solve that conjecture; while trying, it found an argument raising a lower bound for the proportion of zeta zeros on the critical line from 41.6% to 67.2%.

The mathematical move

The result combines work by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh, which made Montgomery-style techniques available without assuming the Riemann hypothesis, with a quadratic-form argument from Bombieri. Anthropic’s short technical synopsis says Claude put the relevant functions in a space carrying Weil’s quadratic form, treated the subspaces arising from zeros on and off the line together, and applied first- and second-moment information to bound the form’s rank. The claimed novelty is not a new theory of zeta zeros but the decision to retain positive and negative definiteness in one non-diagonal quadratic form, making the earlier ingredients yield the stronger constant.

This is precisely the kind of claim for which rediscovery matters. Claude’s own process downloaded and searched 54 arXiv papers, and several subagents reviewed, attacked, and independently re-derived the argument. That is useful internal evidence, not an independent literature review. The post credits Anthropic mathematicians Levent Alpöge and Ralph Furman with studying and validating the work and producing the concise informal note; Brian Conrey and Dan Goldston examined the paper on short notice. It does not claim that the result has passed conventional peer review.

Discovery as a search system

The path is more informative than the successful endpoint. Across two Claude Code sessions, the model emitted 31 million output tokens. An initial pass generated 650 ideas and found none that worked. A second, day-and-a-half pass coordinated about 60 subagents, ran 2,400 shell commands, wrote hundreds of Python scripts, and performed thousands of numerical checks against known zeta zeros. The footnote accounts for the division of labor: two agents developed the key ideas, thirteen fed them candidate ideas, thirty unsuccessful agents explored alternatives, thirteen validated arguments, and two helped draft the paper. Jarred Sumner, the non-mathematician who initiated the run, reports that his intervention after the initial prompt was mostly encouragement.

The successful result was therefore an unintended byproduct of a failed, open-ended target, not the output of a fixed evaluator optimizing a known quantity. Against jiang2026, this is stronger evidence for a research-agent workflow than another saturated-benchmark score, while still matching that survey’s caution: the model composed an extensive human literature into an insight-plus-short-argument result, exactly the regime in which its field map says current systems work best. Nothing in the post shows the model inventing a new concept or reorganizing the subject.

Evidence and trust

Eric Easley worked with Claude to formalize the result in Lean, and Anthropic says the released formalization passes Lean’s comparator. That provides a checker-backed proof object rather than asking readers to accept a corporate claim. It still leaves the two trust layers already visible in this library: demoura2026 concerns whether the checker is sound, while jiang2026 concerns whether the checked statement faithfully captures the mathematical claim. The post links the artifacts but does not report an independent audit of that correspondence.

Read beside openai2026, the evidentiary stack is similar — model-generated argument, human-prepared exposition, Lean formalization — but the discovery shapes differ. OpenAI reported ten targeted solutions and foregrounded their certificates; Anthropic foregrounds a long, mostly failed search whose useful result was adjacent to the prompt. The latter supplies a more concrete account of orchestration cost and negative search, but it still comes from the organization whose unreleased model is being showcased.

Assessment