The Riemann hypothesis says that every nontrivial zero of the Riemann zeta function lies on one vertical line in the complex plane, the critical line. It has been open since 1859. Short of proving it, number theorists have tried to show that at least some fixed share of the zeros lies on that line, and the record share had crept up over eighty years. This post, published on 10 August 2026 under Anthropic's Science label, reports that an unreleased research version of Claude raised that unconditional lower bound "from 41.6% to 67.2%" (opening section). The result was not the task it was given: a staff member had asked it to "take a real stab" at the Riemann hypothesis itself, which it did not prove. The post states plainly that Anthropic does not expect the techniques to lead to a proof of the hypothesis.
The post is a short announcement (about 1,400 words). The mathematics lives in documents it links: the paper, credited to Claude as author; a five-page informal note by Anthropic mathematicians; a 95-page account of how the argument was found, written by Claude from the session logs; a 116-page annotated transcript of two subagent runs; and a Lean formalization. On 13 August two Anthropic mathematicians, Levent Alpöge and Ralph Furman, posted the paper to arXiv under their names (arXiv:2608.13637; v2 on 19 August 2026), stating on the first page that "The mathematical argument in this paper was discovered and written by Claude" and "The listed authors verified the proof and take responsibility for its content." This note follows arXiv v2 for the mathematics and the post for the account of the work. The post was revised: its changelog records a new version of the paper on 13 August, and its text now credits Aryan alongside Baluyot and co-authors, where the first Wayback Machine capture (10 August 2026, 17:47 UTC) names only the latter.
The problem and the prior record
Zeros on the critical line
The zeros of the zeta function control fine detail in the distribution of the prime numbers. Counting them in a window of heights between T and 2T gives a total N(T, 2T) that grows like (T/2π) log T. A proportion result shows that at least some fraction c of those zeros, for all large T, lie on the critical line. A zero is simple if it has multiplicity one; results often count zeros that are both simple and on the line, a stronger property.
The paper's introduction (arXiv v2, Section 1.3) traces the line of results: Selberg showed in 1942 that a positive proportion of zeros lie on the line, Levinson's method gave one third in 1974, Conrey more than two fifths in 1989, and later work brought the record past 5/12 (Pratt, Robles, Zaharescu and Zeindler, 2020). The paper gives 5/12 as the previous record for simple zeros on the line, and the post's "41.6%" is 5/12 (41.67%) cut to one decimal. The 2020 paper's own figures are more precise and differ by kind of zero: more than 0.417293 of the zeros on the line and at least 0.407511 simple and on the line (Pratt and co-authors, arXiv:1802.10521 v3, Section 1). Lamzouri's later paper (below) cites these two figures. Under the assumption that the Riemann hypothesis is true, much more was known: Montgomery's 1973 work on the pair correlation of zeros (how the spacings between zeros are distributed) gave at least two thirds simple zeros, and Montgomery and Taylor's optimized version gave 0.6725 (Section 1.1).
The recent unconditional work the argument builds on
Montgomery's argument needed the Riemann hypothesis at one point: to read a sum over zeros as a sum of positive terms over real numbers. Its other half, an average over primes, is unconditional. The post and paper credit recent papers that made more of the argument work without the hypothesis: Aryan (2022); Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh (2024 and 2025); and Goldston and Suriajaya (2025 and 2026). According to the paper (Sections 1.3 and 7.1), the last showed that the two-thirds figures would follow if all zeros lay within o(1/log T) of the line, and named the termwise positivity that fails off the line as the remaining obstacle, which the paper says its theorem removes. The post also credits a 2000 paper by Bombieri, which, according to the paper (Sections 1.3 and 7.1), observed that the number of negative directions of finite pieces of Weil's quadratic form counts the pairs of zeros off the line.
What the paper states
The main theorem (arXiv v2, Theorem A and abstract) states, unconditionally, that as T grows:
- at least two thirds of the zeros between heights T and 2T, counted with multiplicity, are simple and lie on the critical line;
- at least five sixths of them are distinct;
- with the Montgomery-Taylor choice of test window in place of the simplest one, the two constants become 0.67250 and 0.83625.
The two constants are 2 − R and (3 − R)/2, where R is a constant of the window: R = 4/3 for the simplest window gives 2/3 and 5/6, and R = 1/2 + (1/√2)cot(1/√2) ≈ 1.32750 gives 0.67250 and 0.83625 (Section 1.2; this note recomputed both). The paper gives the previous unconditional records as 5/12 for simple zeros on the line and 0.6603 for distinct zeros (Section 1.1). Lamzouri's paper (Section 1) gives other prior records: the 2020 figures above, and more than 70% distinct zeros, which it credits to Ki and Lee (2012); this note did not check the Ki and Lee paper. The post's headline "67.2%" is the 0.6725 constant. The theorems are asymptotic: they count zeros with heights between T and 2T as T grows (and, by Remark 6.1, between 0 and T). Theorem B extends the same statements to the L-functions of primitive Dirichlet characters, and a remark (Section 7.3, Remark 7.1) gives analogous bounds for the zeros of the derivative of the completed zeta function.
How the argument works, in outline
The paper's own outline (Section 1.2) has three ingredients. In plain terms:
- A finite matrix from Weil's form. Weil's explicit formula links sums over zeros to sums over primes and defines a Hermitian form on test functions; that form is positive on all test functions exactly when the Riemann hypothesis holds. The argument restricts the form to roughly N(T, 2T) shifted copies of one fixed window function spread across the height range, giving a large real symmetric matrix G.
- What each kind of zero contributes. Up to a small error, each zero on the line contributes a rank-one positive piece to G, and each pair of zeros placed symmetrically off the line contributes a block with one positive and one negative direction (Sylvester's law of inertia is the tool here).
- Two computed quantities and one inequality. The trace of G is fixed by the count of zeros (it is asymptotically N), and the sum of squares of its entries is computed from the prime side, using the unconditional second-moment results of Aryan and of Baluyot and co-authors (Section 1.2, items Z and P). A new rank-trace inequality for Hermitian matrices (Lemma 3.2, proved with von Neumann's trace inequality) then bounds from below how much of G can come from simple zeros on the line.
The post's technical paragraph names as the decisive step "The courage to treat the entire space" (section "Claude's finding"). The paper says that the rank-trace inequality is the matrix form of Montgomery's integer step, m² ≥ 2m − 1 (Section 1.2), and that it is not aware of a previous use of rank and positive index together with a second-moment evaluation (Sections 1.3 and 7.1).
What the paper says the result is not
Section 1.4 of the paper, headed "What the results are not," lists the limits. The theorems are lower bounds only: the remaining third of the zeros is not shown to be off the line, only not reached by the method. The analytic inputs are insensitive to a vanishing share (o(N)) of zeros off the line, and they also hold for Davenport-Heilbronn and Epstein zeta functions, for which the analogue of the Riemann hypothesis is false. Given only the quantities the method uses, the inequality is sharp. Among windows the Montgomery-Taylor constant is optimal, the ceiling over a broader class of such certificates is about 0.682 (a bound that rests on numerical enclosures computed outside Lean), and going further would need pair-correlation information beyond Fourier support 1 (Sections 1.1, 1.4 and 7.2). The post agrees: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis."
How the result was found
The post's account
According to the post (section "Claude's methodology"), the work took two sessions in Claude Code and 31 million output tokens. Jarred Sumner, an Anthropic staff member who is not a mathematician, posed the task. In the first session Claude tried 650 ideas, none of which worked. In the second it spent a day and a half coordinating about 60 subagents, which ran 2,400 shell commands, wrote hundreds of Python scripts, checked numerically against known zeta zeros and refereed one another. Footnote 1 sorts the 60: two developed the key ideas, 13 contributed ideas, 30 tried and failed, 13 checked arguments and two wrote the first draft. Subagents then reviewed the proofs, searched for counterexamples, downloaded 54 arXiv papers to check novelty and re-proved the result from scratch, and Claude recommended that a human number theorist check it.
The longer accounts
arXiv v2 adds Appendix B, which the authors say was included because "transparency on these points has been requested by a number of mathematicians who have read the paper." It places the first session in late July 2026 (over six hundred candidate approaches, of which 106 survived internal review and none as a partial proof) and the second over three calendar days in early August (Appendix B.1). The 95-page companion account, written by Claude from the logs, carries a caveat on its first page that "no text in it was written or edited by a human" and that its intermediate mathematics "has not been independently verified." By that account (opening chapter, "The campaign from the coordinator's seat", Sections 1 to 11):
- The decisive run was not aimed at proportions of zeros. A subagent asked to bound the number of negative directions of the Weil form from above found that route empty, and instead used the positive directions to argue that at least half of the zeros lie on the line. The orchestrating model's first reaction was that "my prior is that it's wrong," and it sent the claim to three blind hostile referees, each assigned one possible failure point. arXiv v2 (Appendix B.3) times this run at three and a half hours.
- After the human asked to "Push it to 2/3", a second subagent found the rank-trace inequality, against the steer of the brief it was given (Sections 9 and 11). Two further referees checked it (Section 9). Referee X put the lemma at "at least 0.99" but "2/3 unconditionally" overall at about 0.4; referee Y gave 0.93 for the narrower claim that two thirds follows given the one-half theorem's analytic input. The two numbers answer different questions.
- The first circulated draft claimed two thirds of zeros on the line but only one half simple and three quarters distinct; later model referees noticed that the same lemma gives two thirds simple and five sixths distinct, which became Theorem A (arXiv v2, Appendix B.4).
Who checked the result
- Model referees. All checking during the sessions was by other Claude instances. The 95-page account's Figure 1 labels the final draft "refereed only by other Claude instances," and the orchestrating model's verdict on it was "The next reader should be a person" (opening chapter, Section 10).
- Anthropic mathematicians. Alpöge and Furman "read the argument in detail, checked it independently, condensed it" and placed it in the literature (arXiv v2, Appendix B.5). They are the arXiv authors and take responsibility for its content. The informal note is their concise restatement for experts.
- Outside experts. The post thanks Brian Conrey and Dan Goldston, "two experts in this area, who generously examined the paper on short notice." The paper's acknowledgments thank them "for carefully reading the manuscript and for their comments." Neither document reports what they concluded, and this note found no public statement by either.
- Lean formalization. A Lean 4 formalization of Theorems A and B, orchestrated by Eric Easley, is in the repository anthropics/zeta-23-lean at tag v1.0 (commit 3635e74, 10 August 2026). Its README states that the top-level theorems have no hypotheses, that the repository declares no axioms and that
#print axiomsreports only Lean's three standard axioms; the analytic inputs (explicit formula, zero counting, prime estimates) are proved there, not assumed, and the statements are restated for the comparator tool against Mathlib's definitions alone. The ceiling result alone depends on enclosures computed outside Lean. This note did not build the formalization. - An independent proof. Youness Lamzouri (Université de Lorraine) posted arXiv:2609.02882 (v1 on 2 September, v2 on 8 September 2026), "a new, conceptually simpler, unconditional proof" of the same constants, 0.6725 for simple zeros on the line and 0.8362 for distinct zeros, replacing the matrix framework with "a single Hilbert space inequality" and Montgomery's pair-correlation theorem in its unconditional form (abstract and Section 1). Lamzouri calls the Claude result "the breakthrough result recently announced" and the original argument "technically intricate" with a mechanism "not immediately transparent." His paper adds two further estimates not in the Claude paper, and its Appendix A reports a formal certificate generated by a separate AI system, AxiomProver, under stated assumptions (the unconditional pair-correlation lemma and the zero-counting formula are assumed, not proved).
- Later work. By their arXiv abstracts (not read in full for this note), later preprints apply the method or Lamzouri's version to short intervals (Wang, arXiv:2609.07918) and to runs of moderate gaps between zeros (Angón, Ross, Tang and Zhao, arXiv:2610.22927). A claimed improvement of the 0.6725 constant by about 7 × 10⁻⁸ (Wang, arXiv:2609.24167 v1, 21 September 2026) was withdrawn by its author on 6 October 2026. A guest post on Kevin Buzzard's Xena blog by seven formalization researchers (16 October 2026) names "Anthropic's work on the density of Riemann zeta zeroes" among lab results "done well," on which "the community subsequently built."
Neither the arXiv paper nor Lamzouri's had appeared in a peer-reviewed journal as of 13 November 2026, as far as this note found.
The prompts and instructions used
The post quotes the starting prompt ("take a real stab") and says the human's input during the second session was "mostly limited to sending Claude messages of encouragement (mostly variants of 'keep going' or 'believe in yourself')," which "seems to have helped Claude overcome some initial skepticism." Footnote 2 says a prompt with similar encouragement was used to help Claude disprove the Jacobian conjecture, linking a social-media post; this note did not examine that claim.
The companion documents describe the human's role in more detail. arXiv v2 (Appendix B.1) gives the session's opening instruction as including "take a big leap of faith in your capabilities", and quotes the model declining: "That's not a confidence problem I can fix by believing harder." Its closing remarks (Appendix B.6) say encouragement "was useful in that it licensed breadth; it was not useful as a substitute for verification." The 95-page account lists further one-line prompts that set direction, including "Let's do all rungs now", a request for "more ideas and directions", and "Push it to 2/3", and records that three of the first wave's research directions grew from seeds the human offered (opening chapter, Sections 4 to 8). By the account's Appendix A ("Literature along the way"), the human typed "Push it to 2/3" three minutes after the orchestrating model had itself named two thirds as "a real, well-posed target". The account calls the human's part "thin in content and decisive in direction" (Section 11). Most steering text was written by the orchestrating model: briefs to subagents stating the target, the objects, a forecast and a demand to test every idea on functions for which the Riemann hypothesis is known to fail. No prompt or brief was varied under control, so no document measures what the encouragement or the wording changed.
Limitations and open questions
- One run, reported by the developer. The result came from one campaign, and the account of how it was found was written by Claude from its own logs and is marked as not independently verified. The earlier session's many failed approaches are described only in summary.
- Human verification is internal plus two named readers whose assessment is not reported. The Lean formalization and Lamzouri's independent proof are the strongest checks of the theorem's truth.
- Novelty rests on a literature search that found no prior use of the method; Lamzouri treats the result as new.
- The baselines differ between sources. The paper's prior records (5/12 and 0.6603) differ from those in the 2020 paper and in Lamzouri's survey; the new constants exceed all of them.
- The method has a ceiling of about 0.682 by the paper's own analysis (Section 7.2), and cannot address the Riemann hypothesis.
- Unreleased model. The model was an internal research version, so the run cannot be repeated by outside groups with the same system.
Research questions
- What did the outside readers conclude? Conrey and Goldston read the manuscript on short notice. A published comment from them, or a referee report if the paper is submitted to a journal, would show how specialists outside Anthropic assess the argument.
- How much did the orchestrator's briefs shape the two key steps? The 95-page account separates what the briefs supplied from what the subagents found. Comparing briefs and outputs across all sixty launches, rather than the two successes, would show how often the decisive idea came from the brief.
- Did encouragement change the outcome, or only the length of the attempt? The post suggests encouragement helped overcome skepticism. Repeating comparable attempts with and without encouraging prompts, holding the task and budget fixed, would separate the two.
- How reliable were the model referees? The account records referees that found a real gap and others that gave confidence numbers for differently scoped claims (0.4 and 0.93). Seeding arguments with known errors and measuring how often blind model referees catch them would calibrate that check.
- How does the rank-trace argument relate to Lamzouri's Hilbert-space proof? Lamzouri says both reduce to a quadratic form over zeros and the same extremal problem. A comparison by specialists would show whether either route extends further, for example to the bounds in Lamzouri's two additional estimates.
- What share of the 650 first-session ideas were already known? The companion account sorts the 106 survivors into restated theorems, statements equivalent to the hypothesis, finite checks and near-tautologies. A similar audit of the full set would measure how much of an agent's broad search reproduces known material.
How it connects
- Formalizing Fermat's Last Theorem cites this result as an example of AI-driven work that "produced novel mathematics," in contrast with its own work on verifying a known proof. Both projects rely on Lean formalization checked with the comparator tool.
- Introducing our Science Blog frames the Science posts and points to AI helping mathematicians find new proofs; this is one such case. The connection is the note's own.
- Vibe physics: The AI grad student and Long-running Claude for scientific computing describe other long, lightly supervised research runs. This note draws the comparison; the post does not.
- Claude computes a nine-loop amplitude in N=4 super-Yang-Mills is a later Science post with a different verification arrangement, an outside expert. The connection is the note's own.
References
- Anthropic (2026, August 10; changelog update August 13). Claude has improved on a longstanding lower bound for the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis. https://www.anthropic.com/research/riemann-zeta record
- Alpöge, L., and Furman, R. (2026). More than two thirds of the zeta zeros are simple and on the critical line. arXiv:2608.13637, v2 (19 August 2026). https://arxiv.org/abs/2608.13637. Argument credited to Claude. Locations in this note refer to v2. record
- Claude (2026). More than two thirds of the zeros of the Riemann zeta function are simple and on the critical line. Paper linked from the post, dated 11 August 2026 (the revision the post's changelog describes). https://www-cdn.anthropic.com/95c246936988e43127bc6b2ceb7077c1dad2d68e.pdf. An earlier 35-page version dated 10 August 2026, titled "...lie on the critical line", was also read: https://www-cdn.anthropic.com/564f962e60643842f5fcb4a17c9dbc8f608f1c37.pdf record
- Anthropic (2026). How the two-thirds argument was found: two agent runs and their literature. 95 pages, written by Claude from the session logs. https://www-cdn.anthropic.com/d7f3ecf1d01392d887f8bc974ca187e2a121b1ed.pdf record
- Anthropic (2026). Informal note, 67% of the zeroes are on the line. https://www-cdn.anthropic.com/23455459f8832d06bb175cc0f88d019aed962ef8.pdf record
- Anthropic (2026). Transcripts of Claude sub-agents E2 and E2-pairs, typeset and annotated (116 pages; not read in full for this note). https://www-cdn.anthropic.com/8a0d1add3c637b858a9a181e98c40e9548c3f44f.pdf record
- Lean formalization: https://github.com/anthropics/zeta-23-lean, tag v1.0 (commit 3635e74, 10 August 2026); README read at that tag.
- Lamzouri, Y. (2026). A new proof that more than 2/3 of the zeros of the Riemann zeta function are simple and on the critical line. arXiv:2609.02882, v2 (8 September 2026). https://arxiv.org/abs/2609.02882 record
- Pratt, K., Robles, N., Zaharescu, A., and Zeindler, D. (2020). More than five-twelfths of the zeros of ζ are on the critical line. Res. Math. Sci. 7, Paper 2. arXiv:1802.10521 v3 read for the figures cited. https://arxiv.org/abs/1802.10521 record
- Avigad, J., Buzzard, K., Commelin, J., Degenne, R., Morrison, K., Nash, O., and Topaz, A. (2026, October 16). How to formalize all of mathematics. Guest post, Xena (blog). https://xenaproject.wordpress.com/2026/10/16/how-to-formalize-all-of-mathematics/ record · 2 pages
- Wang, B. (2026). Proportions of the non-trivial zeros of the Riemann zeta function. arXiv:2609.24167, v1 (21 September 2026), withdrawn in v2 (6 October 2026); abstract and withdrawal notice read. https://arxiv.org/abs/2609.24167 record