Claude's Riemann Result Is Real, and the Workflow Is the Story

Fifty years of expert effort moved one of number theory's most-watched dials by about eight points. An AI just moved it twenty-five more in a day and a half — and this time, the receipts are unusually good.

Anthropic says an unreleased research version of Claude has raised the proven lower bound on the fraction of Riemann zeta zeros satisfying the Riemann hypothesis from 41.6% to 67.2%. Not the hypothesis itself — that would require 100%, and Anthropic is explicit that Claude's techniques won't get there. But as partial progress goes, this is a serious jump on a problem with a serious pedigree: Levinson got the bound to one-third in 1974, Brian Conrey ground it past 40% in 1989, and the record has crept upward by fractions of a percent ever since, sitting near 5/12 after the work of Pratt, Robles, Zaharescu, and Zeindler.

Every AI-does-math announcement since 2024 has deserved a skeptical squint, and this one got mine. It mostly survived. Here's why, and what the workflow behind it means for the rest of us.

What actually happened

The setup is almost comically mundane. Jarred Sumner — an Anthropic staffer and not a mathematician — asked Claude, running inside Claude Code, to take a real stab at the Riemann hypothesis. His contribution thereafter was, by Anthropic's own account, mostly variants of "keep going" and "believe in yourself." The first session generated 650 ideas. All failed.

The second session is where it gets interesting. Claude spent roughly a day and a half coordinating about 60 subagents: two developed the core mathematical ideas, thirteen contributed supporting ones, thirty tried new directions and failed, thirteen acted as validators, and two drafted the paper. Along the way the swarm ran 2,400 shell commands, wrote hundreds of Python scripts, and hammered the emerging claims with thousands of numerical checks against known zeta zeros. Total output: 31 million tokens across the two sessions.

The mathematical move itself was recombination, not invention. Claude took a recent line of work by Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh — which made Montgomery's 1973 pair-correlation machinery usable without assuming the hypothesis you're trying to bound — and spliced it with a 2000 paper of Bombieri's, via an inequality on the rank of a quadratic form. Two known results, one non-obvious bridge, a 25.6-point jump.

The verification stack is the real headline

We've been burned before. Last October, OpenAI researchers briefly claimed GPT-5 had solved several open Erdős problems; the "solutions" turned out to be literature search — the model had found existing published proofs the problem list hadn't indexed. The episode was retracted within days and became the canonical example of why you don't take an AI lab's word for a math result.

Anthropic clearly studied that failure, because the verification here is layered the way you'd want. Numerically: thousands of checks against known zeros, plus subagents explicitly tasked with re-proving results independently and hunting counterexamples. Formally: Claude and staffer Eric Easley produced a machine-checked formalization in Lean, published at anthropics/zeta-23-lean for anyone to compile. Socially: two Anthropic mathematicians, Levent Alpöge and Ralph Furman, worked through the paper, and two genuine external authorities — Conrey, who set the 40% record himself, and Dan Goldston, an author of the very work Claude built on — examined it. Furman surfaced in the Hacker News thread to say he and Alpöge took "mathematical accountability" as something like highly interested referees, with formal journal publication still to come.

That last clause matters: this has not yet survived anonymous peer review, and Lean formalizations can hide problems at the statement level even when every proof step checks. But the distance between "trust our benchmark" and "here's a Lean repo, and the guy who held the previous record read the paper" is enormous. This is the most independently checkable AI mathematics claim any lab has shipped.

What developers should take from it

The transferable lesson isn't about zeta functions. It's that the architecture that produced this result is the same one you can build today: a long-horizon agent loop with cheap generation on one side and a merciless verifier on the other. Claude's 650 dead ideas cost nothing because numerical checks and Lean killed the bad ones before a human ever looked. That generate-and-verify shape is exactly why LLMs are already decent at code (the compiler and test suite are the oracle) and still dangerous at freeform prose arguments (no oracle).

So the practical policy for trusting LLM math writes itself. Derivations that arrive with a checkable artifact — executable numerics, property tests, a Lean or Rocq proof that compiles — can be trusted to the extent the artifact covers the claim. Derivations that arrive as confident LaTeX prose are conjecture, whoever generated them. If you're using models for anything quantitative in production, budget for the verifier, not just the tokens; that's where this result says the leverage is.

The other takeaway is about literature scale. No human can hold all of modern analytic number theory in working memory; a model demonstrably can hold enough of it to notice that two papers written decades apart compose. Expect more results of exactly this shape — frontier-of-the-known recombination — in any field with a deep, formalizable literature. That's not "AI creativity," and the HN skeptics are right that the Riemann hypothesis itself likely needs something qualitatively different. It's something narrower and, for working scientists, arguably more useful: exhaustive search over the space of connections experts haven't had time to try.

The caveats you should carry around

You cannot reproduce this with any model you can rent. It was an unreleased research version, and Anthropic makes no claim that shipping Claude models can do this. We also don't know the denominator — how many runs like this quietly produced nothing. One published 36-hour success tells you the capability exists, not its hit rate; the 650-failure first session hints the rate is low. And the "believe in yourself" prompting detail, charming as it is, is a reminder of how unscientific the elicitation side of this still is.

But grade it honestly: a new bound on a 167-year-old problem, a compiling formal proof, and review by the field's actual record-holders. That's not hype. That's the first AI math result I'd cite without holding my nose — and a working blueprint for how to make machine-generated reasoning trustworthy in your own stack: never ship the generation without the verifier.