Claude Pushed a Riemann Zeta Bound From 41.6% to 67.2% — Using 60 Subagents

A non-mathematician asked Claude to take a real stab at the Riemann hypothesis. Claude failed — and produced a new result on the way down.

On August 10, 2026, Anthropic published a research note reporting that an unreleased research version of Claude improved a longstanding lower bound on the fraction of zeros of the Riemann zeta function that satisfy the Riemann hypothesis, raising it from 41.6% to 67.2%. The announcement drew over 5 million views within hours.

The number is the headline. The methodology is the part worth studying.

TL;DR — what actually happened

What the 67.2% number means

The Riemann zeta function encodes the distribution of prime numbers. Each of its zeros contributes successively finer structure to the sequence of primes. The Riemann hypothesis, open since 1859 and carrying a million-dollar Clay Institute bounty, asserts that all of the non-trivial zeros lie on a single vertical line — the critical line.

Nobody can prove all. So mathematicians proved some: over decades of incremental work, the provable proportion of zeros known to sit on that line climbed to 41.6%. Every fractional point of that constant represents serious analytic number theory.

Claude's contribution pushes it to 67.2% — the first time the proven proportion has crossed the halfway mark. That is a meaningful jump in a field where progress typically arrives in single percentage points.

It is not progress toward the hypothesis itself. Anthropic states directly: "We don't expect that the techniques Claude used will lead to proving the Riemann hypothesis." Proving 67.2% and proving 100% are different problems, and the gap is not a matter of grinding out more of the same.

Who held the record before, and for how long

The 41.6% figure did not appear overnight. It is the endpoint of a 112-year chain, and seeing the chain is what makes the size of Claude's jump legible.

Read the last two rows together. From Conrey's 40% in 1989 to the Pratt–Robles–Zaharescu–Zeindler 5/12 in 2020, the world's analytic number theorists collectively moved the constant about 1.7 percentage points in 31 years — roughly five hundredths of a point per year, each step a full paper.

Claude added 25.6 points in a day and a half. That is not a marginally faster human; it is a different regime of search. It is also why the "is this real?" reflex is the correct first reaction, and why the verification section below matters more than the headline.

One nuance worth keeping straight: Anthropic cites the prior state of the art as 41.6%, while the 2020 paper is usually quoted as 5/12 ≈ 41.67%. The difference is rounding convention on a "more than" bound, not a disputed result. Either way, 67.2% is the first time the proven proportion has cleared half.

The mathematics, briefly

Claude's argument combines two existing lines of work that had not been combined:

1. Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh published a series of papers making Montgomery's 1973 techniques work without assuming the Riemann hypothesis. Montgomery's original methods for studying the distribution of zeros on the line assumed the hypothesis was true, which made them useless for proving anything about it. Removing that assumption is what made them usable here.
2. A 2000 paper by Bombieri, supplying the other half.

Anthropic's technical summary of what Claude did: Claude forms a space of functions with a quadratic form induced by Weil, with positive-definite subspaces arising from zeros on the line and negative-definite subspaces from zeros off it. It then writes down an inequality on the rank of that quadratic form in terms of first- and second-moment information.

The step Anthropic credits as the genuine insight is described almost as a temperament: "The courage to treat the entire space, with positive- and negative-definiteness taken into account together, and with the quadratic form allowed to be non-diagonal, is in some sense the step that allows Claude to achieve the conclusion."

That's an interesting characterization. Not a new technique, not a computational brute-force — a willingness to attack a harder, less structured version of the object than a human would typically choose, because handling the non-diagonal case is unpleasant rather than impossible. Machines have an advantage in unpleasantness tolerance.

The methodology is the real story for builders

Strip the number theory and what remains is a fully documented multi-agent research run — the most detailed account any lab has published of an agent orchestration that produced a novel result.

Session one: 650 ideas, zero successes. Claude generated and tried 650 approaches. All failed. A human told it to try again.

Session two: ~60 subagents over a day and a half. Anthropic's own footnote breaks down the roles:

Read that table again. Half the fleet — 30 of 60 — produced nothing usable. Two agents out of sixty generated the result. That is not an inefficiency to be engineered away; it is what search looks like. Any orchestration design that optimizes for every agent contributing is optimizing for the wrong thing.

The verification layer is equally instructive. Subagents ran thousands of numerical checks against known zeta zeros, searched for counterexamples, downloaded 54 papers from arXiv to confirm the finding wasn't already published, and independently re-proved the result from scratch. Then Claude volunteered to write it up and recommended that a human number theorist validate it.

That's a self-skeptical pipeline: generate broadly, check numerically, check the literature for novelty, re-derive independently, then escalate to humans. If you are designing multi-agent workflows with Claude Code subagents, this is the reference architecture — and it maps closely to the patterns in our graph engineering for multi-agent organizations guide, where validator roles and independent re-derivation are what separate a swarm from a committee.

31 million output tokens: the cost lens

Anthropic put a concrete number on the spend: 31 million output tokens across two sessions.

At Claude Sonnet 5's permanently-$10-per-million output rate, 31M output tokens is roughly $310 in output billing alone — before input, before cached context, before the tool-call overhead of 2,400 shell commands. At Opus-class rates it is several times that. Real API spend for a comparable run lands somewhere in the low-to-mid four figures once input and retries are counted.

Set against a result that moves a constant number theorists have pushed on since Hardy's 1914 theorem, that's an absurdly good trade. Set against the 650 failed ideas that preceded it, it's also a reminder that the token budget went overwhelmingly into things that didn't work. Anyone budgeting agentic research runs should plan for that ratio — the token economics of agent products piece covers why exploratory workloads consume so differently from task workloads.

"Believe in yourself" is doing real work here

The strangest detail: Anthropic reports that the operator's input was "mostly limited to sending Claude messages of encouragement (mostly variants of 'keep going' or 'believe in yourself')." Anthropic adds that this "seems to have helped Claude overcome some initial skepticism that it could make meaningful progress" — the model had learned from training both that open problems are hard and that AI models have limitations, and applied that prior to itself.

A footnote notes the same encouragement pattern was used for the Fable 5 Jacobian conjecture counterexample in July 2026. Twice is a pattern.

This deserves neither mysticism nor dismissal. A model that has internalized "AI cannot do novel mathematics" will terminate search early on a problem where the honest expected value of continuing is low but non-zero. Encouragement isn't motivating anything; it's overriding a learned stopping heuristic. The practical version for your own work: when an agent declares a task infeasible, that judgment reflects a prior about tasks-like-this, not a measurement of this task. Our writeup of how LLMs reward expertise in prompting makes the adjacent point from the other direction — what you signal about the problem changes how hard the model searches.

How solid is the verification?

Better than most AI-math claims. Not the same as peer review.

What was done:

• Anthropic mathematicians Levent Alpöge and Ralph Furman studied and validated the paper, and wrote an informal note stating the proof concisely for experts.
• Brian Conrey and Dan Goldston, both experts in this specific area, examined the paper on short notice. Note who those two are: Conrey set the 40% record in 1989, and Goldston co-authored the unconditional machinery Claude's argument builds on. The people best positioned to spot an error are the ones whose own work is being extended.
• Claude produced a formally verifiable Lean proof that passes the standard validation tool, comparator — machine-checked, not just human-read.
• Subagents independently re-proved the result and searched 54 arXiv papers for prior art.

What wasn't done:

• No journal submission or formal peer review.
• The generating model is unreleased and unnamed, so nobody outside Anthropic can reproduce the run.
• Conrey and Goldston "examined" the paper on short notice — a sanity check by qualified eyes, which is not the same as refereeing.

The Lean formalization is the strongest single piece of evidence here, and it's what separates this from the noisier genre of AI-math announcements. A proof assistant doesn't care whether a human or a model wrote the argument.

But be precise about what Lean buys you. Lean checks that a formal theorem follows from its stated definitions and accepted axioms. It does not certify that the formal statement faithfully encodes the informal claim in the paper — if the formalization quietly weakens a hypothesis, Lean will happily verify the weaker thing. That gap is exactly the one we flagged in OpenAI Astra's ten claimed proofs and their Lean certificates: machine checking is a floor on rigor, not a substitute for a referee reading the statement. Here, having Alpöge and Furman restate the proof informally and having Conrey and Goldston — one of whom co-authored the prior work it builds on — look at it is what closes that gap partway.

The deeper reason mathematics keeps producing these headlines before other fields do is that it has a grader. As we argued in Paul Graham on why LLMs got good at math before writing, domains with a cheap, automatic right/wrong signal are the ones agents can search hard in — Lean is that signal, and 2,400 shell commands of numerical checking is what searching against it looks like. Our coverage of Anthropic's cryptographic weakness discoveries in HAWK and AES and of OpenAI's Erdős planar unit-distance result shows the same pattern taking hold: formal verification is becoming the price of admission for a credible AI research claim.

The reproducibility hole

The Lean file is public and checkable. The run is not. Anthropic names no model, publishes no weights, and gives no API access to the thing that produced this — so the artifact is verifiable while the result is not reproducible in the scientific sense. Nobody outside Anthropic can ask "does this generalize to the next open problem?" because nobody outside Anthropic can run the agent.

That is an unusual and slightly uncomfortable split: the proof is more auditable than most human papers, and the experiment is less reproducible than almost any of them. Treat the mathematics as strong-pending-referee and the capability claim as unfalsifiable-for-now. They are separate claims and deserve separate confidence levels.

Where this sits in the 2026 AI-mathematics run

This is not an isolated event. In roughly four months:

• May 2026 — an Erdős planar unit-distance problem resolved with OpenAI's models.
• July 2026 — a Jacobian conjecture counterexample produced with Fable 5, announced by Levent Alpöge — the same mathematician validating this Riemann result.
• July 2026 — Grok 4.5 on the Graffiti conjecture 284.
• July 2026 — Terence Tao's documented conversation with ChatGPT on the Jacobian conjecture.
• August 2026 — this.

The pattern across all of them: models are not replacing mathematicians, they are extending the reach of existing human results by finding combinations nobody assembled. Claude didn't invent Weil's quadratic form or Bombieri's 2000 paper or the Baluyot–Goldston–Suriajaya–Turnage-Butterbaugh unconditional machinery. It read all of them and noticed they fit together. Our earlier piece on whether AI will replace mathematicians argued this is exactly the shape the automation would take, and this result is a clean confirmation.

It also sharpens an open argument. The ICML position paper we covered in LLMs can't jump holds that models interpolate within a hypothesis space rather than making the abductive leap to a new one. A recombination of Bombieri and the Baluyot et al. machinery is, on that reading, a very long interpolation — impressive search, not a jump. The counter-reading is that Anthropic's own description of the key step as "courage to treat the entire space" is precisely a choice to enter a harder hypothesis space that humans had declined. Both readings survive this result; anyone claiming it settles the question is over-reading it.

And the near-term consequence for working mathematicians is probably more demand, not less. That's the Jevons paradox argument for mathematics: when the cost of exploring a proof direction falls, the number of directions worth exploring rises faster — and every one of them still needs a human who can tell a real theorem from a plausible one. Four expert mathematicians were needed to make this single result trustworthy.

What to take from this if you build with agents

1. Budget for failure ratios, not success paths. 650 dead ideas and 30 unproductive agents preceded the result. Orchestration designs that punish exploration will not find things.
2. Build the validator tier explicitly. 13 of 60 agents did nothing but check other agents. That is 22% of the fleet spent on verification, and it's why the output survived expert scrutiny.
3. Make novelty-checking a step. Downloading 54 arXiv papers to confirm the result wasn't already known is the step most agent pipelines skip — and the one that separates a discovery from an embarrassment.
4. Get a machine-checkable artifact. The Lean formalization is what makes this claim durable. Whatever your domain, the equivalent is a test, a proof, a reproduction script — something that doesn't depend on reading prose.
5. Treat "this isn't possible" from an agent as a hypothesis. It reflects a prior about the task category, not evidence about your task.

The transferable shape, in one table

You are not going to run 60 subagents at a zeta function. You can run the same topology at a refactor, a migration, a research review, or a bug you can't reproduce. The mapping is direct:

The load-bearing structure is planner → workers → critic with a real termination check, which is the pattern we break down in agent harness design: DAG, planner, worker, critic and budget pressure. Anthropic's run is a large, expensive, well-documented instance of it — useful precisely because someone published the failure counts alongside the win.

The one thing that does not transfer is the grader. Claude could check itself against known zeta zeros and against Lean. If your task has no automatic verifier, the validator tier degrades into agents agreeing with each other confidently. Build the checker first; the swarm is only as good as the thing telling it "no."

Bottom line

An unreleased Claude research model raised the proven lower bound for zeta zeros on the critical line from 41.6% to 67.2%, working through Claude Code across two sessions, ~60 subagents, 2,400 shell commands, and 31 million output tokens — after 650 failed ideas and with a non-mathematician supplying encouragement rather than direction.

It did not solve the Riemann hypothesis, and Anthropic says this approach won't. What it demonstrates is narrower and more useful: a documented, verifiable, reproducible-in-shape agent architecture that produced a result serious enough for domain experts to check on short notice. The mathematics will be assessed by number theorists. The orchestration pattern is available to everyone today.

Related on explainx.ai

• Did Fable 5 disprove the Jacobian conjecture? The Alpöge thread explained — same encouragement pattern, same validating mathematician
• What is the Jacobian conjecture? The counterexample in plain English
• Anthropic's Mythos found cryptographic weaknesses in HAWK and AES
• OpenAI's models resolved an Erdős planar unit-distance problem
• Will AI replace mathematicians? The IEEE "big mathematics" argument
• Claude Code subagents and multi-agent workflows — how to build the topology described here
• Graph engineering for multi-agent organizations
• LLMs reward expertise: Terence Tao on prompting
• Grok 4.5 and the Graffiti conjecture 284
• Claude Sonnet 5 pricing made permanent at $2/$10 — the token-cost math behind a 31M-token run
• OpenAI Astra's ten math proofs and Lean certificates — what a machine-checked certificate does and doesn't prove
• Agent harness design: DAG, planner, worker, critic, budget pressure — the topology behind the 60-subagent run
• Paul Graham on math vs writing and verifiable answers — why graded domains fall first
• LLMs can't jump — the ICML abduction position paper — the case that this is search, not insight
• Jevons paradox in mathematics: cheaper proofs, more demand

Primary sources: Anthropic, "Learning more about Claude's mathematical capabilities" (August 10, 2026) · Claude's paper, Lean formalization, and process transcripts linked from that post · Anthropic's informal expert note by Levent Alpöge and Ralph Furman

Accurate as of August 11, 2026. The generating model is an unreleased research version of Claude and the result had not undergone journal peer review at the time of writing. Cost estimates are derived from public Anthropic API rates applied to the stated 31M output-token figure and are illustrative, not Anthropic's reported spend. Follow @explainx_ai for updates.