Claude’s Zeta Breakthrough: 67% on the Critical Line

It starts the way a lot of good math stories do: with an impossible challenge.

On August 10, 2026, Anthropic shared an account of an internal “real stab” at the Riemann hypothesis—one of the most famous unsolved problems in math. Claude didn’t crack the hypothesis itself. But while reaching for it, it produced something that does count: a new improvement to a longstanding lower bound on how many zeros of the Riemann zeta function lie on the critical line.

This post walks through the ideas in that result at a beginner-friendly pace, while still respecting the technical core. What makes this story especially interesting isn’t “AI can do math.” It’s that a sophisticated number-theory strategy—built from decades of human research—was turned into a new, provable inequality.

The Riemann zeta function, in plain language

The Riemann zeta function, usually written as (\zeta(s)), is a complex-valued function that packages information about prime numbers.

• The input (s) is complex: (s = \sigma + it), where (\sigma) is the real part and (t) is the imaginary part.
• A zero of (\zeta(s)) is a value of (s) where (\zeta(s) = 0).
• The nontrivial zeros are the zeros in the “critical strip,” the region where the real part (\sigma) lies between 0 and 1.

Now the big geometric picture: in the complex plane, there’s a special vertical line called the critical line, defined by (\sigma = 1/2). The Riemann hypothesis claims that every nontrivial zero lies on that line.

So mathematicians often talk about a different, more modest target: even if we can’t prove every zero lands exactly on (1/2), what minimum fraction can we guarantee?

That’s the number Claude’s work improved: the known lower bound increased from 41.6% to 67.2%.

If you’re hearing these percentages and thinking, “That sounds abstract,” it is. But it’s also concrete: it means “among many zeros up to height (T), the proportion that sit on the critical line can’t be smaller than 67.2% (in the limiting sense used in analytic number theory).”

Why “fraction of zeros” is a real math goal

When people say “lower bound for the fraction of zeros on the critical line,” they’re mixing a few technical notions:

1. A proportion: not a single zero, but a density-like statement about many zeros as (t) grows.
2. A lower bound: a guarantee that this proportion cannot drop below a certain value.
3. A limiting argument: these bounds are understood as (T \to \infty), using the analytic number theory machinery for asymptotics.

This is the kind of theorem you can’t test on a laptop by checking one hundred zeros. It’s a statement about the “shape of the spectrum” of zeta zeros—how they accumulate and distribute.

And here’s where the story gets subtle: improving those bounds usually isn’t about getting a better numerical approximation. It’s about building the right inequality that forces zeros to behave in a constrained way.

From zeros to quadratic forms: the main bridge

Claude’s method (as described in Anthropic’s write-up) lands on an inequality involving a quadratic form.

A quadratic form is a function that takes a vector (think: coordinates in a vector space) and returns a number based on a rule of the form “sum of products of coordinates.” In matrix language, a quadratic form looks like (x^T A x), where (A) is a matrix.

The key idea is to design this quadratic form so that:

• It interacts with the zeta zeros through a carefully chosen analytic framework.
• It separates “good behavior” from “bad behavior” using the sign structure of the form.

That sign structure is expressed through positive-definite and negative-definite subspaces.

• A subspace is a smaller vector space living inside a larger one.
• A quadratic form is positive-definite on a subspace if plugging any nonzero vector from that subspace makes the quadratic form come out positive.
• It’s negative-definite on a subspace if the same setup makes the value negative.

So, informally, the proof strategy tries to split the world of candidate functions into regions where the quadratic form has consistent sign. Then an inequality about the rank of the quadratic form—rank meaning “how many genuinely independent directions the matrix/form has”—turns those sign constraints into a statement about where zeros can be.

This is the moment where the story feels magical. How can a quadratic form tell us where zeta zeros live?

The answer is that the quadratic form is engineered to encode zeta information—so when the rank/sign constraints force certain components to exist, they translate back into “at least this many zeros must be on the critical line.”

Claude’s “surprising step” wasn’t inventing zeta theory from scratch. It was treating the entire function space in a unified way, accounting for both positive- and negative-definite parts together, and allowing a non-diagonal quadratic form rather than simplifying it into a form that loses power.

The prior work it builds on

Number theory has a long memory. This result is described as drawing heavily on a line of research that adapts Montgomery-style techniques without assuming the Riemann hypothesis.

• Montgomery’s pair correlation ideas are about how zeros “pair up” statistically, and they were originally used in contexts that assumed the hypothesis.
• Later work created versions of these ideas that operate without that assumption, enabling lower-bound improvements for zeros on the critical line.

On top of that, Claude also leans on an earlier contribution by Enrico Bombieri, connected to what you might call an “equivalence-style” framework around the Riemann hypothesis.

You don’t need the full history to appreciate the payoff: Claude’s proof turns those tools into a sharper inequality.

What Claude actually did (the practical engineering side)

The math is the headline. But the backstage process matters too, especially for readers curious about how an AI can traverse a technical proof landscape.

In Anthropic’s account, an unreleased research version of Claude produced the new bound over two sessions using 31 million output tokens. The internal workflow involved multiple stages:

• An initial wave of attempts: Claude generated and tried 650 ideas, none of which worked.
• A second push: after a prompt to try again, Claude coordinated roughly 60 subagents for about a day and a half.
• Computation and validation: those subagents ran 2,400 shell commands, wrote hundreds of Python scripts, performed numerical checks against known zeta zeros, and acted as mutual reviewers.
• Proof auditing: it also searched to ensure the result wasn’t already known, including downloading 54 arXiv papers.

One small detail that reads like a human story: the prompting that preceded the successful run was largely encouragement.

That sounds psychological, not mathematical. But in a multi-agent, multi-attempt search process, persistence can matter as much as raw cleverness.

Why formal proof matters: from “convincing” to “verifiable”

Even when the mathematics looks right, formal verification is a different category of confidence.

In this case, Claude also produced a Lean formalization—Lean being a proof assistant where theorems are checked by a logical kernel.

A proof in Lean isn’t “an explanation.” It’s a program-like object whose correctness is verified step-by-step according to the rules of the logic.

Anthropic reports that the formal result passed a standard validation pipeline (using the comparator tool). That’s significant because it addresses the most common failure mode for AI-generated math: an argument that sounds plausible but contains a hidden gap.

Also, before trusting the result broadly, Anthropic had mathematicians examine how the new argument relates to prior work and assess the structure of the proof.

The real meaning of “67.2%”

It’s tempting to interpret 67.2% as “close to solving the Riemann hypothesis.” But the Riemann hypothesis is an all-or-nothing statement.

A lower bound like 67.2% says: in the large-scale limit, you can’t have too many zeros off the critical line. It does not say the off-line zeros disappear.

Still, improvements here are not cosmetic. Each increment comes with a new inequality and often a new technical perspective—new ways of controlling moments, transforms, and quadratic forms that other researchers can reuse.

So the best way to read this result is as a proof-of-process and a proof-of-possibility: AI didn’t just “estimate.” It navigated a rigorous framework, produced a stronger theorem, and then subjected it to verification.

Closing thought: progress that wasn’t the goal

The Riemann hypothesis remains unproven. But the story of Claude’s “related win” highlights a pattern we’re only beginning to understand: sometimes the shortest path to a better theorem is not the path you originally asked for.

Claude chased a conjecture. Instead, it sharpened the tools around it—turning old analytic number theory into a new rank inequality that forces zeros to behave better than anyone previously could guarantee.

That’s not a guarantee that the Riemann hypothesis is next. But it is a vivid sign that model-powered exploration is starting to touch the kinds of constraints that real mathematical progress depends on: engineered representations, carefully controlled inequalities, and proofs that don’t merely persuade—they verify.