---
title: "Study 4: twelve AI coding agents, four rounds, 37% fewer tokens · Cork"
description: "Twelve Claude Code agents at once on one repo, four rounds with Cork and four without, same tasks and model: 13.14M vs 20.87M tokens in total, 37% fewer (29 to 44% per round), 47 of 48 tasks correct vs 46."
canonical: "https://corkcli.com/benchmarks/study-4-twelve-agents-four-rounds"
last_updated: "2026-11-14"
---

# Twelve agents, four rounds: 37% fewer tokens with Cork

Twelve Claude Code agents launched at once on one repository, each with its own task, all pushing to the same origin. We ran it four times with Cork and four times without: same tasks, same prompts, same model. Over the four rounds the agents with Cork used 13.14M tokens and the agents without it used 20.87M. Every round saved between 29% and 44%.

37%
:   fewer tokens, four rounds

30%
:   fewer turns (585 vs 832)

27%
:   lower cost ($8.05 vs $11.00)

47/48
:   tasks correct, vs 46/48 without

## Every round

Tokens for all twelve agents in each round, bars to scale.

Columns: with Cork, without

Round 1 · 44% fewer

3.27M

5.82M

Round 2 · 29% fewer

3.43M

4.83M

Round 3 · 36% fewer

3.36M

5.26M

Round 4 · 38% fewer

3.09M

4.95M

## The numbers

Tokens are input, cache writes, cache reads and output, as the provider reported them for each agent. Cost is the model's list price for that usage. Correct means the task's hidden acceptance tests pass on the final shared origin.

| Round | With tokens | Cost | Correct | Without tokens | Cost | Correct |
| --- | --- | --- | --- | --- | --- | --- |
| Round 1 | 3,265,685 | $1.99 | 12/12 | 5,823,679 | $3.01 | 12/12 |
| Round 2 | 3,430,418 | $2.13 | 11/12 | 4,833,900 | $2.55 | 12/12 |
| Round 3 | 3,359,939 | $2.00 | 12/12 | 5,263,310 | $2.76 | 10/12 |
| Round 4 | 3,088,854 | $1.93 | 12/12 | 4,947,115 | $2.68 | 12/12 |
| All four | 13,144,896 | $8.05 | 47/48 | 20,868,004 | $11.00 | 46/48 |

## How it was run

- **The work.** Study 3's hardest tier: twelve tasks in a 654-file Python repository, several of them vague tickets, with decoys and one task that depends on another agent's work. Each agent got its own clone and pushed to one shared origin.
- **With Cork.** Cork's paid plan, set up the way a new user sets it up: one setup command installed Cork's hooks for the machine, and each agent also had Cork's MCP server. It ran on a staging copy of Cork's production servers, on the same build.
- **Without.** The same agents, prompts and flags with no Cork.
- **Both.** Claude Sonnet 5, all twelve agents launched at once. Any agent whose work had not reached the origin when it stopped got one identical wake-up message, in both groups.

## What this study does not show

- All twelve agents ran under one person's account. Study 3 split them across four people and measured 47% on its hardest tier with an earlier model.
- Agents finished in about a minute each, so these rounds say little about wall-clock time.
- One task failed once with Cork (round 2) and two failed once without it (round 3). Neither group was correct every time.
- In dollars the gap is smaller than in tokens (27% vs 37%), because much of each session is billed at the cache-read rate.

[All studies](https://corkcli.com/benchmarks)[Install Cork](https://corkcli.com/install)

This page in HTML: https://corkcli.com/benchmarks/study-4-twelve-agents-four-rounds