60 Subagents Moved a Century-Old Maths Bound
Ignore the maths for a moment. The orchestration numbers are the most useful thing published about multi-agent work this year.
Founder, Automation Squad ·
The short answer
Anthropic reported on August 10, 2026 that an unreleased research version of Claude improved a longstanding lower bound on the fraction of Riemann zeta zeros satisfying the hypothesis, from 41.6% to 67.2%. The run used two Claude Code sessions, roughly 60 subagents, 31 million output tokens and 2,400 shell commands over about 36 hours. Anthropic states plainly it is not a proof.
What the run actually looked like, and what transfers to normal work
The maths does not transfer. The shape of the run does — and there is very little published data on what a large multi-agent run costs in practice.
Take the ratio, not the headline
Sixty subagents and 31 million output tokens to move one result over 36 hours. That is the first public datapoint on what heavy parallel orchestration actually costs, and it is worth having in mind next time someone proposes throwing agents at a problem.
Note that it was two sessions, not sixty independent runs
The subagents sat inside a small number of coordinating sessions. Parallelism was used for exploration underneath a coherent thread, rather than as sixty separate attempts hoping one worked.
Notice what made it verifiable
The output was checkable — a mathematical bound, plus a Lean 4 formalisation. Heavy parallel exploration works best where the result can be independently verified, because that is what filters confident nonsense out of sixty parallel attempts.
Apply the pattern where verification is cheap
In ordinary work that means: exhaustive search over a codebase, testing many hypotheses against data you hold, generating many candidates against a rubric you can apply. Not: writing something subjective sixty ways.
Read Anthropic's caveats as part of the finding
Not a proof, not conventionally peer reviewed, not reproducible end to end because the model is unidentified. A company publishing limits that specific alongside a headline result is doing something worth encouraging.
| Measure | Figure |
|---|---|
| Bound improved | 41.6% → 67.2% |
| Claude Code sessions | 2 |
| Subagents | ~60 |
| Output tokens | 31 million |
| Shell commands | 2,400 |
| Wall-clock time | ~36 hours |
| Reviewed by | 2 in-house mathematicians, plus Brian Conrey and Dan Goldston |
| Is it a proof? | No — and not peer reviewed |
| Reproducible end to end? | No — the model is unidentified |
Take it with you
ANTHROPIC'S RIEMANN RUN — the orchestration numbers
Aug 10, 2026
Source: https://automationsquad.com/news/claude-riemann-subagents/
THE RUN
Bound improved ......... 41.6% -> 67.2%
Claude Code sessions ... 2
Subagents .............. ~60
Output tokens .......... 31,000,000
Shell commands ......... 2,400
Wall clock ............. ~36 hours
Reviewed by ............ 2 in-house mathematicians +
Brian Conrey, Dan Goldston
ANTHROPIC'S OWN CAVEATS
Not a proof
Not conventionally peer reviewed
Not reproducible end to end (the model is unidentified)
WHAT TRANSFERS
Not the maths. The SHAPE.
- 60 subagents inside 2 COORDINATING sessions, not 60 separate runs
- the output was CHECKABLE (a bound + a Lean 4 formalisation)
- heavy parallel exploration works where verification is cheap
USE THE PATTERN WHERE VERIFICATION IS CHEAP
[x] exhaustive search over a codebase
[x] many hypotheses tested against data you hold
[x] many candidates scored against a rubric you can apply
[ ] writing something subjective sixty ways <- noAnthropic published on August 10, 2026 that an unreleased research version of Claude had improved a longstanding lower bound on the fraction of Riemann zeta zeros satisfying the hypothesis, moving it from 41.6% to 67.2%.
The facts: the run took two Claude Code sessions, roughly 60 subagents, 31 million output tokens and 2,400 shell commands over approximately 36 hours, alongside thousands of numerical checks. Anthropic states clearly that this is not a proof, that it has not undergone conventional peer review, and that it cannot be reproduced end to end because the model is unidentified. Two in-house mathematicians reviewed it, along with Brian Conrey and Dan Goldston. Anthropic released a formal paper, a Lean 4 formalisation, an informal note and full process transcripts.
Automation Squad's take: the mathematics is not the transferable part and most coverage stopped there. What is genuinely useful is that somebody finally published the cost of a large multi-agent run — 60 subagents, 31 million output tokens, 36 hours, one result. Anyone who has been asked whether to "just run more agents at it" now has a reference point. Two details matter more than the totals. The subagents ran underneath two coordinating sessions rather than as sixty independent attempts, which is a different architecture from the fan-out people usually imagine. And the output was checkable, with a Lean 4 formalisation to verify against — heavy parallel exploration is only useful where you can filter the results, because sixty parallel attempts at something unverifiable produce sixty confident answers and no way to choose.
Run this now: look at your own work for a task where generating many candidates is cheap and checking them is also cheap — searching a codebase exhaustively, testing a list of hypotheses against data you already hold, producing many drafts against a rubric you can actually apply. That is where this pattern earns its cost. If the checking step is expensive or subjective, parallelism will multiply your output and not your confidence, and the 31-million-token figure is a useful thing to remember before starting.
Questions people are asking
- Did Claude prove the Riemann hypothesis?
- No. Anthropic is explicit that this is not a proof. An unreleased research version improved a lower bound on the fraction of zeta zeros satisfying the hypothesis, from 41.6% to 67.2% — a real result in its own right, and not the same thing as proving the hypothesis.
- How many subagents did the run use?
- Roughly 60, across two Claude Code sessions, consuming 31 million output tokens and 2,400 shell commands over approximately 36 hours.
- Can the result be reproduced?
- Not end to end, because the model used is an unreleased research version and has not been identified. Anthropic released a formal paper, a Lean 4 formalisation, an informal note and full process transcripts, and had the work reviewed by two in-house mathematicians plus Brian Conrey and Dan Goldston.
Sources
Further reading
- Anthropic — Riemann zeta research — Includes the formal paper, the Lean 4 formalisation and the full process transcripts.
Last checked August 10, 2026 against the primary sources above, by Automation Squad Research. Spot an error? [email protected].