---
title: "Study 5: twelve AI coding agents, 53% fewer tokens with Cork · Cork"
description: "Twelve Claude Code agents at once on one repo, with Cork's paid plan and without, same tasks and model: 5.48M vs 11.77M tokens, 53% fewer, 51% fewer turns and 24% less wall-clock time."
canonical: "https://corkcli.com/benchmarks/study-5-twelve-agents-53-percent"
last_updated: "2026-11-14"
---

# Twelve agents, one repo: 53% 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, once with Cork's paid plan and once without: same tasks, same prompts, same model. The agents with Cork used 5.48M tokens; the agents without it used 11.77M. They also took 51% fewer turns and finished 24% sooner.

53%
:   fewer tokens

51%
:   fewer turns (152 vs 312)

44%
:   lower cost ($2.44 vs $4.32)

24%
:   less wall-clock time (284s vs 373s)

## The numbers

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

Columns: with Cork, without

5.48M

11.77M

| Twelve agents | With Cork | Without |
| --- | --- | --- |
| Tokens | 5,475,538 | 11,770,652 |
| Cost | $2.44 | $4.32 |
| Turns | 152 | 312 |
| Wall-clock time | 284s | 373s |
| Tasks correct | 11/12 | 12/12 |

## How it was run

- **The work.** Study 3's hardest tier: twelve tasks in a 654-file Python repository 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, with prompts matched to code by meaning: Cork's setup command, its hooks, and its MCP server for every agent, on a staging copy of Cork's production servers.
- **Without.** The same agents, prompts and flags with no Cork.
- **Both.** Claude Sonnet 5, all twelve agents launched at once under one account. Any agent whose work had not reached the origin when it stopped got one identical wake-up message.

## What this study does not show

- One round each way. Four more rounds two days later, with Cork set up at the machine level ([Study 4](https://corkcli.com/benchmarks/study-4-twelve-agents-four-rounds)) measured 37% across the rounds.
- One task (refunds by day) failed with Cork and passed without it: its spec allows two readings of the dates it returns.
- In dollars the gap is smaller than in tokens (44% vs 53%), 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-5-twelve-agents-53-percent