The result is serious enough to deserve attention, because Anthropic has published more than a marketing summary. Its accompanying 35-page paper, “More than Two Thirds of the Zeros of the Riemann Zeta Function Lie on the Critical Line,” lays out the theorem, names its mathematical inputs, identifies the prior record, and includes a public Lean 4 formalization. But the claim remains newly released work rather than a peer-reviewed result, and Anthropic has not said which Claude model produced it, when—if ever—it will be available to customers, or whether users will be able to reproduce the research workflow in Claude Code.
For IT professionals watching AI agents move from code completion toward autonomous research, the important development is not a vague claim that an LLM “did math.” It is that Anthropic describes a system that searched literature, ran numerical tests, developed competing proof approaches, assigned review work to subagents, generated a formal proof artifact, and surfaced a result that human mathematicians then assessed. The practical lesson is that agentic tooling is beginning to look less like an answer generator and more like a high-throughput research environment—although only in a carefully provisioned, highly resourced internal setting for now.
The 67.25% figure is a lower bound, not partial completion of RH
The Riemann hypothesis says every nontrivial zero of the zeta function has real part (1/2), placing every one of them on the critical line. It remains unproved. A proof that 67.25% of zeros are on that line leaves open the possibility that the remainder are also on it, but it provides no evidence sufficient to conclude that they are.
Anthropic’s public post rounded the new number to 67.2%, while Claude’s paper gives the optimized constant as 0.6725, or 67.25%. The old benchmark is likewise usually expressed as more than (5/12), which equals 41.666…%, rather than the post’s rounded 41.6%. Those are presentation differences, not conflicting records, but the precision matters in number theory: the result is an asymptotic lower bound, not a count of zeros that were computationally checked.
The paper claims three linked improvements. It proves that at least 67.25% of zeros are on the critical line; that at least 67.25% are simple zeros on that line; and that at least 83.625% are distinct zeros. The critical-line portion is the headline result because it surpasses the longstanding unconditional 41.66% record associated with the Levinson-method line of work.
Earlier work by Siegfred Baluyot, Daniel Goldston, Ade Irma Suriajaya, and Caroline Turnage-Butterbaugh had already removed the Riemann-hypothesis assumption from an important part of Montgomery’s pair-correlation machinery. Their 2024 Acta Arithmetica paper, however, did not establish that a comparable proportion of zeros were on the line without further assumptions. Anthropic’s paper identifies the unresolved obstacle as the loss of a positivity argument once off-line zeros are admitted.
Claude’s proposed advance is to handle that obstacle using a finite-dimensional compression of Weil’s Hermitian form and a linear-algebra argument about its rank, trace, and positive and negative directions. In plainer terms, zeros on the critical line and symmetric pairs off it contribute mathematically different signatures to a matrix-like object. The proof claims that the first two moments supplied by the prime side of Montgomery’s method are enough to force a large minimum number of on-line zeros without assuming RH.
That is a genuine conceptual difference from merely optimizing the previous mollifier calculations. It also explains why the result has attracted immediate scrutiny: a short route around a limitation that stood for decades is exactly the sort of result that must survive close reading by number theorists.
The paper itself sets a hard limit on what this method can establish
Anthropic deserves credit for placing the strongest caveat in the technical paper rather than burying it in promotional language. The paper says the argument has “no bearing” on the Riemann hypothesis in either direction and only certifies a lower bound. It does not show that the unaccounted-for roughly one-third of zeros lie off the critical line; they are simply beyond what this certificate proves.
More importantly, the authors state that the mathematical ingredients used here are insensitive to a small number of off-line zeros. The paper notes that similar inputs can hold for related functions whose Riemann-hypothesis analogue is false. That sharply limits the usual leap from “large proportion” to “therefore close to a full proof.”
There is also a mathematical ceiling in the method as presented. Claude’s paper says that, using only its particular bandwidth-one pair-correlation information and the associated configuration-by-configuration certificate, the approach cannot prove a simple-zero proportion above roughly 68.185%. Reaching substantially larger constants would require information about prime pairs beyond the range currently available through the same machinery.
So the 67.25% result is large in historical context, but it is not a direct runway to 100%. Anthropic says as much in its announcement, stating that it does not expect the techniques used here to prove RH. That is a more credible framing than treating a related theorem as a near-solution to a Millennium Prize problem.
Lean formalization is stronger evidence than numerical checking—but it is not external peer review
The most consequential detail for technically minded readers is the Lean 4 repository. Anthropic’s repository says it contains a complete,
sorry-free formalization of its Theorems A through E, with no declared axioms beyond Lean’s standard logical foundations. It also says the formalization includes the analytic machinery needed for the result: Weil’s explicit formula, zero-counting results, estimates involving the gamma function, prime-sum estimates, and the generalized Hilbert inequality used in the proof.
If that description holds under independent inspection, it is much stronger than saying a model ran numerical checks against known zeta zeros. Numerical checks can catch obvious bad conjectures but cannot prove an asymptotic theorem. A proof assistant can verify that every formal step follows from encoded definitions and prior theorems.
The crucial qualification is what has been formalized. A Lean checker verifies the theorem statement placed in the source tree, not whether a press release’s prose accurately captures its mathematical significance. In this case, Anthropic’s repository says its theorems are formulated directly using Mathlib’s definition of the Riemann zeta function and carry no extra hypotheses, which directly addresses the most common concern about a formal result being proved only conditionally or with a key assumption hidden in the setup.
Still, the formalization is hosted by Anthropic and was released alongside the company’s own paper. Anthropic says mathematicians Levent Alpöge and Ralph Furman studied the result, and says Brian Conrey and Daniel Goldston examined the paper on short notice. Those are meaningful checks, especially given Goldston’s direct connection to the precursor work, but Anthropic has not published independent referee reports from Conrey or Goldston, and no peer-reviewed journal publication exists as of August 11, 2026.
That leaves the correct status as: a public proof claim with unusually substantial supporting artifacts, including mechanized verification, awaiting broad independent mathematical review.
Claude Code’s research workflow was a compute-heavy agent operation
Anthropic says the result emerged during two Claude Code sessions that produced 31 million output tokens. According to the company, staff member Jarred Sumner—described as a non-mathematician—initially asked the model to make a serious attempt at RH. Claude generated and discarded 650 ideas before a second attempt coordinated around 60 subagents over about a day and a half.
The stated workflow included 2,400 shell commands, hundreds of Python scripts, thousands of numerical checks, literature searches across 54 arXiv papers, adversarial proof review, counterexample searches, and independent re-proofs by separate subagents. Anthropic’s own breakdown is revealing: only two of the 60 agents were responsible for developing the key mathematical ideas, while others contributed suggestions, attempted alternative approaches, validated arguments, or helped produce the paper.
This looks much closer to an engineering research pipeline than the popular image of a chatbot composing an answer in one pass. The useful capability is orchestration: persistent exploration, tool use, division of labor, self-criticism, and the ability to turn an apparent finding into artifacts that specialists can inspect.
But it also makes clear why this should not yet be read as a feature announcement for Claude Code users. Anthropic has disclosed output-token volume but not the model identity, hardware budget, token pricing equivalent, context configuration, tool permissions, prompting framework, or exact agent harness needed to reproduce the work. “Unreleased research version” is doing significant work in the announcement.
For organizations evaluating AI agents, that gap matters. A workflow requiring dozens of coordinated agents, unrestricted paper retrieval, local scripting, extended runtime, and millions of generated tokens belongs in a controlled research environment with logging, sandboxing, cost controls, and human review. It is far removed from asking a standard enterprise assistant to investigate a problem during an ordinary workday.
Anthropic has published the proof, but not a product roadmap
There is no indication that the research version of Claude is available through Claude.ai, the Claude API, or Claude Code. Anthropic has announced neither a model name nor a release date, and it has not committed to offering the multi-agent research configuration described in the paper. Readers should therefore treat this as evidence about Anthropic’s internal frontier capability, not a new customer-facing mathematical-research product.
The immediate consequence is narrower but still significant. Anthropic has put a specific, falsifiable mathematical claim into public view, together with a conventional paper, an expert-facing note, process transcripts, and a formal verification repository. The result will now stand or fall on independent attempts to inspect the argument and rebuild the Lean project.
If the proof withstands that examination, the bigger story will not be that Claude came 67.25% of the way to solving the Riemann hypothesis. It will be that an unreleased agent system found a new route through established mathematics, documented it well enough for formal checking, and turned a failed open-problem attempt into a theorem mathematicians can actually evaluate.