Key Notes
- An unreleased research version of Claude raised a longstanding lower bound on the fraction of Riemann zeta zeros proven to lie on the critical line from 41.6% to 67.2%, the largest single jump in this figure's history.
- It did not solve or make progress toward the full Riemann hypothesis; Anthropic says the technique is unlikely to lead to a proof, and the result classifies a population of zeros rather than measuring progress toward 100%.
- Claude ran as a swarm of about 60 subagents in Claude Code over two sessions, using 31 million output tokens and 2,400 shell commands, after 650 failed ideas; the human prompter, a non-mathematician, mostly offered encouragement like "keep going." The result was formally verified in Lean, validated by two Anthropic mathematicians, and reviewed by external number theorists Brian Conrey and Dan Goldston; the Lean proof is public.
- It has not yet passed conventional peer review and cannot be fully reproduced because the model is unreleased.
Anthropic said on August 10 that an unreleased research version of Claude improved a longstanding result connected to the Riemann hypothesis, one of mathematics’ most famous unsolved problems. Asked to “take a real stab” at the hypothesis by a non-mathematician staff member, Claude did not solve it, but along the way it raised a decades-old lower bound: the proven minimum fraction of the Riemann zeta function’s nontrivial zeros that lie on the “critical line” the hypothesis predicts.
Claude increased that bound from 41.6% to 67.2%, drawing heavily on prior work by mathematicians including a series of recent papers by Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh, and a 2000 paper by Bombieri. Anthropic’s own mathematicians validated the result, Claude produced a machine-checkable proof in the Lean proof assistant, and two external number theorists, Brian Conrey and Dan Goldston, reviewed the paper.
The scale of what was and was not achieved needs to be stated precisely, because online reaction outran the result. This is not a proof of the Riemann hypothesis, which would require showing that 100% of the relevant zeros lie on the line, and it does not mean researchers are “67.2% of the way” there. Nor does it locate the remaining zeros or show that any lie off the line. It is a stronger unconditional guarantee that at least 67.2% of the zeros satisfy the condition, leaving open the possibility that all of them do.
Anthropic itself said it does not expect the techniques Claude used to lead to a full proof. Within those bounds, the jump is genuinely notable: mathematicians had raised this particular bound by less than one percentage point over the prior 37 years, so a 25.6-point improvement is a large single step, which one investor called potentially the most significant analytic number theory advance since the 2013 bounded-prime-gaps result.
The methodology drew as much attention as the result. Claude worked across two sessions in Claude Code, first generating 650 ideas that failed, then coordinating a swarm of about 60 subagents over roughly a day and a half. Together they ran 2,400 shell commands, wrote hundreds of Python scripts, ran thousands of numerical checks against known zeta zeros, downloaded 54 arXiv papers to confirm the finding was novel, and refereed one another’s work. The whole run consumed 31 million output tokens.
The human involved, staff member Jarred Sumner, supplied almost no mathematical direction, and his input was mostly messages of encouragement, variants of “keep going” and “believe in yourself,” which Anthropic says helped Claude overcome initial skepticism that it could make progress. The key mathematical move, per Anthropic, was treating zeros on and off the line as a single unified space rather than analyzing them separately.
Why the Verification Matters More Than the Number
The most important feature of this announcement is not the percentage but the presence of an automatic verifier. Claude could check its own work against known zeta zeros and, decisively, formalize the proof in Lean, which mechanically checks every logical step rather than relying on human review.
That is what separates this from the confident-sounding but wrong “proofs” language models often produce, and it is why external experts could validate it quickly. The broader lesson practitioners drew is instructive: a swarm of AI subagents is powerful precisely when a rigorous checker can tell it “no,” and far less trustworthy on problems that lack one, where agents can simply agree with each other. The result is strong evidence that AI can now extend real mathematical ideas in a verifiable way, on a problem with a built-in grader.
The Caveats and the Bigger Picture
Several caveats keep this in proportion. The result has not undergone conventional peer review, the process by which the mathematics community independently scrutinizes and accepts a claim, and it cannot be reproduced end to end because Anthropic used an unidentified, unreleased model. Anthropic is also an interested party announcing a capability of its own product, so the framing is naturally favorable, even as the underlying math appears to hold up under the external checks performed so far.
The advance fits a pattern of rapid progress in AI mathematical reasoning, following systems from Google DeepMind and OpenAI posting strong results on competition problems, and Anthropic says Claude contributed to other open-problem progress earlier in 2026. The honest reading is that this is a real, verified contribution to analytic number theory produced with heavy human scaffolding and prior human research, not a machine independently cracking a landmark conjecture.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
