Claude Raised the Riemann Zeta-Zero Bound from 41.6% to 67.2%

On August 10, 2026, Anthropic released a research note that quietly reset what people expect an LLM to do inside mathematics. An unreleased research version of Claude — not a shipping model — improved the proven lower bound on the fraction of Riemann zeta zeros that lie on the critical line from 41.6% to 67.2%. It did not prove the Riemann hypothesis. Anthropic is explicit that the techniques Claude used are not expected to get there. What Claude actually did is arguably more instructive for AI builders: it synthesized a chain of recent analytic number theory papers into a new theorem, produced a paper, generated a Lean 4 formalization of its result, and survived review by two in-house mathematicians and two of the world's leading experts.

This is the first AI-generated contribution to number theory serious enough that Anthropic's staff mathematicians Levent Alpöge and Ralph Furman studied and validated it, and that external experts Brian Conrey and Dan Goldston — names every analytic number theorist recognizes — examined the manuscript on short notice. That is an E-E-A-T bar AI research rarely clears.

What the zeta-zero result actually means

The Riemann zeta function ζ(s) is defined for Re(s) > 1 by the infinite sum Σ n⁻ˢ and continued analytically to the whole complex plane. Its nontrivial zeros — the ones in the critical strip 0 < Re(s) < 1 — control how prime numbers are distributed. In 1859, Riemann conjectured that all nontrivial zeros satisfy Re(s) = ½, the critical line. That conjecture, the Riemann hypothesis, remains unproven after more than 165 years and carries a Clay Institute million-dollar bounty.

Because nobody can prove that all zeros are on the line, mathematicians have spent a century proving that a growing fraction of them must be. Levinson passed one-third in 1974 using the mollifier method. Conrey passed two-fifths in 1989. A 2020 paper by Pratt, Robles, Zaharescu and Zeindler pushed past 5/12 ≈ 41.67% — the "41.6%" Anthropic cites as the previous state of the art. Claude's result pushes the proven proportion above two-thirds for the first time ever.

Old bound vs new bound

A worked explanation of the method

Montgomery's 1973 pair-correlation work introduced the machinery for counting zeros on the line, but many of its cleanest consequences assumed the Riemann hypothesis itself. A series of recent papers by Baluyot, Goldston, Suriajaya and Turnage-Butterbaugh (arXiv:2306.04799, arXiv:2501.14545) showed how to make Montgomery's techniques work without that assumption. Claude's contribution was finding that those results, combined with Bombieri's 2000 paper, unlock a bound the community had missed. The combination is the "synthesis."

At a high level, the proof works like this:

1. Zero-counting becomes a positive sum. Using Weil's explicit formula, the distribution of zeros up to height T is connected to sums over primes up to X = (T/2π)^λ, where λ is a band-limit parameter ≤ 1.
2. The obstacle is replaced by linear algebra. Earlier work needed a termwise positivity argument that fails off the critical line. Claude replaces it with a finite compression of Weil's Hermitian form, read via Sylvester's law of inertia — a textbook linear algebra tool, not a new deep estimate.
3. The arithmetic side stays light. The proof uses only Chebyshev–Mertens estimates for prime-power sums and the Montgomery–Vaughan generalized Hilbert inequality. No zero-density estimates, no zero-free regions, no mollifiers.

Theorems A–C prove respectively: at least (2/3 − o(1)) of zeros lie on the critical line; the same fraction are simple zeros on the line; and at least 5/6 of zeros are distinct. An optimized test family recovers exactly the Montgomery–Taylor kernel and yields constants 0.6725, 0.6725 and 0.83625 (Theorem D). The same statements hold for primitive Dirichlet L-functions (Theorem E).

The paper also proves a ceiling: no certificate of this kind, reading only bandwidth-one pair-correlation data, can certify more than 0.68185 simple zeros. Reaching 0.70, 0.80 or 0.90 by the same route requires pair-correlation information on Fourier supports out to roughly 1.04, 1.26 and 1.70 respectively — beyond what is known. That honest self-limitation is a large part of why mathematicians took the result seriously.

What "synthesizing recent papers" means for LLM research workflows

The word that matters is synthesizing. Claude did not derive the result from first principles, and it did not brute-force search. Given the challenge in Claude Code, the model first generated roughly 650 candidate ideas that failed. Then, across two sessions with ~60 subagents, ~2,400 shell commands and hundreds of Python scripts, it connected the Baluyot–Goldston–Suriajaya–Turnage-Butterbaugh program to Bombieri's older work in a combination nobody had published.

The workflow shape is the story. Subagents ran numerical checks against known zeros, searched for counterexamples, downloaded 54 arXiv papers to confirm the result had not already been published, and independently re-proved the result from scratch as a consistency check. Anthropic put the model's own paper, the Lean 4 formalization, and process transcripts online. That is a research pipeline with a paper trail, not a lucky autocomplete.

Anthropic reports the total spend as 31 million output tokens. At current public pricing for reasoning-heavy frontier models (very roughly $15–$40 per million output tokens), the direct API cost lands in the $500–$1,200 range — before internal tooling compute. Even at the high end, that is cheaper than a month of a postdoc's time, and the previous half-dozen percentage points of progress each took years of expert effort. The marginal cost of a candidate theorem is collapsing; the cost of verifying one is now the bottleneck.

How it differs from a formal proof — verification methods

There are three artifacts, and they are not interchangeable:

1. The paper — Claude's manuscript (authored by the model) stating Theorems A–E, with an account of the process in Appendix C.
2. The informal expert note — Alpöge and Furman's concise statement of Claude's argument for human readers.
3. The Lean 4 formalization — Claude produced a formally verifiable proof in Lean (github.com/anthropics/zeta-23-lean). Machine-checked logic means the formal object compiles, but the correspondence between paper and code still requires human scrutiny, and the formalization encodes Claude's chosen argument rather than independent methodology.

The asymmetry is worth internalizing: a model that can produce results is increasingly common, but a model that can produce a Lean 4 artifact and survive adversarial re-derivation by subagents is not. Verification discipline — not raw reasoning power — is the differentiator.

What it signals for AI mathematical research

Three signals, each with a caveat.

• Agent research harnesses work in math. The ~60-subagent, 2,400-shell-command topology is reproducible in shape today with Claude Code, LangGraph, or any orchestration framework. The bottleneck has moved from generation to verification.
• The field advances by combination, not revolution. The single most valuable LLM skill here was finding that paper A + paper B + paper C fit together. That is exactly the skill our scientific literature synthesis workflows have been automating for applied domains.
• Skepticism scales with the claim. Nobody should read "67.2%" as progress toward the full hypothesis; Anthropic states the gap is structural. Benchmark hygiene — pinned dates, pinned model versions, released artifacts — is what made this claim checkable in the first place. The NIST TEVV-Athlon framework argument applies to math capability claims exactly as it does to agent safety claims.

Why This Matters for Developers and Researchers

• Reproduce the harness, not the theorem. Two sessions, ~60 subagents, adversarial re-derivation, Lean output. We shipped a miniature version of this loop at SaaSNext — an agent swarm that reads fresh arXiv submissions in a target domain and proposes lemmas against a pytest harness. The failure rate is brutal: over 90% of proposals are wrong, trivial, or uninteresting. But when you can run 650 hypotheses in a day at sub-dollar marginal cost, the asymmetry still wins. That is the transferable lesson, and it generalizes far beyond number theory.
• Plan for token budgets. A 31M-output-token run is a real line item. Before launching multi-day autonomous research, read our context window economics guide and checkpoint aggressively so a failed branch does not burn a week of budget.
• Buy verification infrastructure early. Whether it is Lean, proof scripts, or property-based test harnesses, the team that verifies fastest wins. Our verifiable AI inference coverage shows the same principle in applied settings.
• Pin benchmarks to versions. This ran on an unreleased research build. It says nothing about shipping Claude, and comparing against our earlier Claude 3.7 Sonnet Extended Thinking vs DeepSeek-R1 reasoning audit, the honest rule is: test the artifact, never the headline.

The honest limits

Anthropic is unambiguous: the pair-correlation-plus-linear-algebra technique cannot prove the Riemann hypothesis. The method has a proven ceiling of 0.68185 for simple zeros, and closing the gap from 67.2% to 100% is structural, not incremental. Claude jumped the bound by 25.6 percentage points in a field where progress usually arrives in single digits — and then explained precisely why it cannot finish the job. For researchers, that combination of breakthrough and honest ceiling is the most valuable artifact of all.

Sources: Anthropic research note (August 10, 2026) · Claude's paper · Lean 4 formalization · Baluyot–Goldston–Suriajaya–Turnage-Butterbaugh (arXiv:2306.04799) · Bombieri (2000).

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Last tested/verified: August 2026 with Anthropic's unreleased research build of Claude running in the Claude Code harness, Lean 4 formalization at github.com/anthropics/zeta-23-lean, and the linked paper PDF dated August 10, 2026. Not journal peer-reviewed as of writing; cost figures are illustrative estimates derived from public API rates, not Anthropic-reported spend.