Claymath.org
An unreleased research version of Claude spent 36 hours autonomously coordinating roughly 60 AI subagents — generating 650 failed ideas before landing on a cross-domain insight that two existing papers could be combined in a way no human mathematician had tried — and emerged with the largest single-step advance on a famous 167-year-old math problem in decades. On August 10, Anthropic’s research post announced that the model had raised the proven lower bound for the fraction of Riemann zeta function zeros lying on the so-called “critical line” from 41.6% to 67.2% — a 25.6-percentage-point jump accomplished in roughly 36 hours.
According to one observer, a partner and researcher at Menlo Ventures, it may be the most significant number-theory advance since 2013, a field that had moved that same lower bound only 0.8 percentage points over the previous 37 years.
The result did not prove the Riemann hypothesis — one of mathematics’ seven Millennium Prize Problems, with a $1 million bounty from the Clay Mathematics Institute for a complete solution. But what the run produced was something specific, verifiable, and already mechanically confirmed: a public Lean 4 proof on GitHub, that any mathematician in the world can check for themselves. The method itself is available for inspection. The AI model that produced it is not.
What the Riemann Hypothesis Actually Is
To understand why a lower bound on zeta zeros matters, a brief foundation. The Riemann zeta function — first fully studied by Bernhard Riemann in an 1859 paper — is a mathematical object whose behavior encodes how prime numbers are distributed among the integers. The function has infinitely many “nontrivial zeros,” points where it equals zero, and the Riemann hypothesis conjecture page describes how every one of those zeros is conjectured to lie on a specific vertical line in the complex plane — the critical line, where the real part of the complex argument equals exactly 1/2.
Nobody has proved or disproved it. Because a complete proof remains out of reach, mathematicians have instead spent decades working on the adjacent question: what minimum fraction of those zeros can we rigorously prove lies on the critical line? G.H. Hardy proved in 1914 that infinitely many zeros are on the line; Levinson showed in 1974 that the fraction is at least 1/3; Conrey improved that to 2/5 in 1989; and a 2018 paper by Pratt, Robles, Zaharescu, and Zeindler established the previous state of art at 41.6%.
Claude’s run raised that floor to 67.2%.
From a Dare to a Day and a Half of Subagents
Jarred Sumner — the creator of the Bun JavaScript runtime, who joined Anthropic when Bun was acquired in December 2025 — gave the model a prompt that a professional mathematician would have found alarming: “Take a real stab at the Riemann hypothesis.” Sumner’s prompt left every mathematical choice to the model. He is not a mathematician.
Claude’s first response was something resembling productive skepticism. Over the initial session, the model generated and evaluated 650 distinct ideas — every one of them a dead end. Sumner prompted it to try again, with input that Anthropic describes as “mostly variants of ‘keep going’ or ‘believe in yourself.'” What followed was a day and a half of autonomous multi-agent work inside Claude Code: approximately 60 subagents operating in parallel, running 2,400 shell commands, writing hundreds of Python scripts, performing thousands of numerical checks against known zeta zeros, and having subagents critique one another’s reasoning. The agents also downloaded and reviewed 54 papers from arXiv to confirm the eventual finding hadn’t already been published. Total compute across both sessions: 31 million output tokens.
The encouragement pattern is not incidental. The model had internalized from its training both that open mathematical problems are very hard and that AI systems have limitations — and had initially applied those priors conservatively to its own search. A similar encouragement approach, Anthropic notes, was used earlier in 2026 to help Claude produce a counterexample to the Jacobian conjecture, another longstanding open problem.
The Specific Cross-Paper Bridge No Human Had Made
The mathematical substance of the result turns on a connection between two bodies of work that no researcher had previously combined.
In 1973, Hugh Montgomery introduced a set of pair correlation techniques for studying the distribution of zeta zeros. Those techniques were powerful — but they assumed the Riemann hypothesis was true, which meant they could not be used to prove results about zeros unconditionally. More recently, Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh published unconditional pair correlation results that removed that assumption, producing unconditional versions of Montgomery’s methods that could support work on raising the lower bound.
Separately, a 2000 paper by Enrico Bombieri — a Fields Medal winner — studied a quadratic form associated with Weil’s explicit formula, the identity linking prime numbers to zeta zeros. Bombieri showed that this quadratic form is positive semidefinite if and only if the Riemann hypothesis holds, and that violations of the hypothesis correspond exactly to negative eigenvalues of the form.
Claude’s insight was to treat the entire function space simultaneously — accounting together for the positive-definite subspace (zeros on the critical line) and the negative-definite subspace (zeros off it), allowing the associated quadratic form to be non-diagonal. By writing down a rank inequality using first- and second-moment information, computed via the dual picture over primes and Hilbert transform control, the model derived a bound that improved on what either body of work could produce alone. In Anthropic’s own description, the result “emerged as the unintended byproduct” of attempting the full hypothesis.
Note that Dan Goldston — one of the external reviewers who examined Claude’s paper — is also a co-author of the Baluyot et al. papers that Claude’s proof relies on. His presence in both capacities strengthens the validation: the researcher who helped produce the upstream machinery reviewed whether the downstream application of it was correct.
Verification: Lean Proof, Four Named Reviewers, No Journal Submission Yet
Two in-house mathematicians at Anthropic — Levent Alpöge and Ralph Furman — examined Claude’s paper and produced an informal note for experts, summarizing the proof concisely. Two external analytic number theory specialists, Brian Conrey and Dan Goldston, reviewed the work on short notice. Claude also worked with Anthropic staff member Eric Easley to produce a Lean 4 formalization of the result — a mechanically verifiable version of the proof that passes the standard validation tool comparator.
This verification package represents an unusual asymmetry in the epistemic landscape of the result. The mathematical claim is more auditable than almost any comparable human-produced paper: the Lean formalization means that anyone with access to the proof assistant can check the logical derivation step by step, without trusting the author’s judgment or reading ability. At the same time, the AI capability claim is less reproducible than almost any other published AI benchmark: the model used is unreleased, with no public weights, no accessible checkpoint, and no API. A researcher who wants to verify that Claude produced this result has to take Anthropic’s word for the process, even as the mathematical output itself is independently checkable. As one independent analysis observed, this puts reviewers in the position of validating a proof without access to the tool that generated the ideas.
As of August 12, the paper has not undergone conventional peer review at a mathematics journal.
What It Doesn’t Mean — and What the 32.8% Tells You
Anthropic is direct on one point: the techniques Claude used are not expected to lead to a complete proof of the Riemann hypothesis. Raising the lower bound to 67.2% establishes that at least two-thirds of zeta zeros are on the critical line. It says nothing about the remaining 32.8% — those zeros may well also be on the line, but the proof does not establish it. The Clay Prize requirements page makes clear it requires a complete proof, not an improved lower bound. The hypothesis itself remains open.
What the result does establish is a verifiable advance on a benchmark that the mathematics community had moved only incrementally over 46 years of combined human effort. The analogy to the bounded prime gap result of 2013 — when Yitang Zhang proved for the first time that infinitely many pairs of primes are separated by at most 70 million, breaking a decades-long standstill — is imperfect but instructive. Both results were notable not just for what they proved, but for demonstrating that a long-stalled problem was still accessible to new techniques.
What the Template Implies Beyond This Result
The methodology of the Riemann zeta run — many autonomous subagents targeting a hard technical problem, numerical verification at scale, Lean formalization as a mechanical audit layer, and named human experts as the closing validation step — is attracting at least as much attention as the result itself.
This is not the first time Claude has surfaced unexpected mathematics in 2026. Earlier this year, Claude Fable 5 contributed to resolving the Jacobian conjecture. Around the same period, Google’s Gemini resolved several open Erdős problems in combinatorics, and OpenAI’s GPT disproved a longstanding discrete geometry conjecture. The Riemann zeta result fits a pattern: frontier AI models finding non-obvious connections between existing bodies of work across subfields, rather than inventing new mathematical concepts from scratch.
As one analysis of the methodology observed, models are not replacing mathematicians; they are extending the reach of existing human results by finding combinations nobody assembled. That framing — extension rather than replacement — fits the Bombieri plus Baluyot architecture precisely. Claude did not invent Weil’s quadratic form, Bombieri’s 2000 analysis, or the unconditional Montgomery machinery. It read all of them and noticed they fit together in a way that yielded a new bound.
Whether that capability holds as problems grow harder is unanswered. The Riemann hypothesis itself, and the other six Millennium Prize Problems, remain open. But for the lower-bound question specifically, the record now stands at 67.2% — set in 36 hours, with a public Lean proof, by an AI whose weights remain out of reach.
Frequently Asked QuestionsDid Claude prove the Riemann hypothesis?
No. Claude raised the proven lower bound on the fraction of Riemann zeta function zeros lying on the critical line — from 41.6% to 67.2%. This is a verifiable mathematical advance on a related, more tractable question. The Riemann hypothesis itself — which predicts that all nontrivial zeros lie on the critical line — remains unproven and undisproven, as it has since 1859. The Clay Mathematics Institute’s $1 million prize for a complete proof has not been claimed.
What exactly is the Lean 4 formalization, and does it prove the result is correct?
Lean 4 proof assistant is an open-source system that mechanically checks whether a system that mechanically checks whether a mathematical argument’s logical steps are valid, using the same type-checking that verifies computer programs. Claude’s Lean formalization of the Riemann zeta result is publicly available on GitHub and passes the standard validation tool comparator. This means the logical derivation can be verified by any researcher with access to the proof assistant — the result does not require trusting Claude’s judgment or Anthropic’s description of the proof. However, as researchers in AI-assisted formal verification note, machine type-checking confirms logical correctness; it does not guarantee that the formalized statement faithfully captures the intended mathematical content. That alignment judgment still requires human expert review, which Conrey, Goldston, Alpöge, and Furman provided.
Can this multi-agent approach be applied to other open problems?
That is the live question the result leaves open. The template — many autonomous subagents, numerical verification at scale, Lean formalization as a mechanical audit layer, named human experts closing the validation — is in principle generalizable. Anthropic used a similar pattern in the Jacobian conjecture work earlier in 2026, and the same subagent-plus-Lean recipe also underpinned Claude Mythos Preview’s cryptanalysis results in July. The honest answer is that no one yet knows which open problems are accessible to this approach, or how reliably it produces results of comparable quality across different domains. What is now established is that it worked once on a benchmark that human researchers had barely moved in 46 years. For a broader read on this agentic research pattern, AI Weekly covers the template’s implications in detail.
Is the result independently reproducible?
The mathematics is independently auditable — via the public Lean proof and the released paper. The AI run itself is not reproducible: the model used is an unreleased research version of Claude, with no public weights, no checkpoint, and no API access. Researchers can verify that the proof is correct; they cannot verify that an AI system will produce a comparable result on a different problem, because the specific model is not available to them. This auditability-reproducibility asymmetry is a novel situation for the mathematics and AI research communities alike, and it is one the community will need to develop norms around as AI-assisted mathematical research becomes more common. One independent review covers reproducibility concerns remain in detail.