# Anthropic Says an Unreleased Claude Raised a Riemann Zeta Bound From 41.6% to 67.2%

The lab reports the model produced a formally verifiable proof, and says it does not expect the technique to lead to a proof of the Riemann hypothesis itself.

- Published: 2026-09-01T06:29:06.347Z
- Canonical: https://polylog.news/ai/2026-09-01/anthropic-says-an-unreleased-claude-raised-a-riemann-zeta-bo
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Anthropic Research](https://www.anthropic.com/research/riemann-zeta)

Anthropic [reported](https://www.anthropic.com/research/riemann-zeta) that an unreleased research version of Claude improved a long-standing lower bound on the fraction of nontrivial zeros of the Riemann zeta function that lie on the critical line, raising it from 41.6 percent to 67.2 percent. The Riemann hypothesis asserts that all such zeros lie on that line. Proving a positive fraction, and then pushing that fraction up, has been a decades-long program in analytic number theory.

Anthropic states that two mathematicians on staff studied and validated the model's paper, and that Claude also produced a formally verifiable proof of the result. A machine-checkable proof matters more than the headline percentage, because it removes the usual failure mode in which a model produces a plausible argument with a subtle gap.

The lab is explicit about the limit of the claim. It does not expect the techniques used here to lead to a proof of the Riemann hypothesis. This is an improvement to a quantitative bound inside an established framework, not a new route to the conjecture.

Two caveats matter here. The model is not released, so no outside group can reproduce the run, and the mathematicians who checked the work are employed by the company that benefits if the claim holds. Independent number theorists working through the argument, and the formal proof passing an outside proof assistant check, are what would move this from a vendor report to an established result.

## What this means

The interesting signal is the verification pipeline, not the theorem. A model that emits a machine-checkable proof alongside its argument converts mathematical output from something requiring expert review into something a computer can accept or reject, which is the same property that made reinforcement learning with verifiable rewards work for code and competition math. If that generalizes to research-level mathematics, labs gain a training signal in a domain where human labeling does not scale, and Anthropic gains a capability argument it can point to without shipping the model. If independent mathematicians find the formal proof does not cover the full claimed result, the story reverts to a vendor demonstration.

## What to watch

- Whether outside number theorists publish an assessment of the argument, which is the difference between a lab claim and an accepted result.
- Whether the formal proof is released in a checkable form so third parties can run it themselves.
- Whether Anthropic ships this research version or folds the technique into a released model, which would tell you if the capability is a training-run artifact or a durable product feature.