---
title: "Benchmarks: token cost of multiple AI coding agents, measured · Cork"
description: "Measured token cost for AI coding agents sharing one repo. In the latest study, agents on Cork used 5.3× fewer tokens per agent than agents searching the repo themselves (240K vs 1,262K), and the cost per agent fell from 545K to 240K as ten agents joined. Every transcript published."
canonical: "https://corkcli.com/benchmarks"
last_updated: "2026-11-14"
---

# What multiple AI coding agents actually cost, measured

Every number here comes from recorded agent sessions, not estimates, and every headline ratio compares two groups of agents from the same run: same task, same model, same day. The only difference is whether the agents shared context through Cork or searched the repository themselves.

5.3×
:   fewer tokens per agent than agents searching the repo themselves, same run

100%
:   tasks landed correctly, every round

10 → 0
:   edits thrown away when four agents share one repo

21/21
:   features correct against the live server

Study 2 · 2026-09-26

## Ten agents with Cork, ten without, one repository

Each agent had to thread a value through every caller of one function in a 300-module Python repo, where the ticket does not say where that value lives. On Cork, the first agent found it and saved the change as a verified recipe; each later agent was handed that recipe by the hooks before it began and finished in four or five turns. Without Cork, every agent started over: 37.2 turns on average. All 20 agents landed the change correctly.

5.3×
:   fewer tokens per agent, same run

240K
:   per agent with Cork, vs 1,262K

545K → 240K
:   cost per agent as ten agents joined

5.5 vs 37.2
:   turns per agent

Legend: with Cork; searching the repo themselves (without)

Running cost per agent

**The more agents share the repo, the less each one costs.** The first agent pays for discovery; every agent after it inherits the answer. Without shared context there is no trend, because every agent starts from nothing.

The whole run is checkable: all 20 session transcripts add up to the exact totals the provider reported, and the server's hash-chained ledger of 828 rows verifies end to end.

[Read Study 2](https://corkcli.com/benchmarks/study-2-ten-agents-one-repo)

## Two ways to count

By Graphify's arithmetic

621×

1,182,646 tokens an agent spent discovering the repository on its own, over the 1,904 tokens of facts a briefing hands it instead. Raw over compact, per query, the way the 71.5× is computed.

vs

What you are billed

5.3×

Study 2: the whole session, per agent, ten agents each way, same task, same model, same run, every one correct. The agent still edits, runs the tests and reads what it must; the facts are the small part. This is the number we publish.

## Studies

Each study is a set of agent runs with its own method and raw numbers. New studies are added as rows; nothing in an earlier study is rewritten.

1. [Study 1

   Sep 18–20, 2026

   Shared context, round by round

   Change a function signature and every caller in a 300-module repo; then a discovery task where the answer must be found first.

   6–10 per round, rounds 7–13 agents · 1,211K → 196K tokens per agent; four concurrent agents with 0 edits thrown away

   6.2×

   fewer tokens per agent, Round 11, same run](https://corkcli.com/benchmarks/study-1-shared-context)
2. [Study 2

   Sep 26, 2026

   Ten agents, one repo, today's build

   The discovery task again, 10 agents with Cork and 10 without, every transcript and ledger row published.

   20 (10 + 10) agents · 1,262K → 240K tokens per agent; cost per agent fell from 545K to 240K as agents joined

   5.3×

   fewer tokens per agent, same run](https://corkcli.com/benchmarks/study-2-ten-agents-one-repo)
3. [Study 3

   Sep 28, 2026

   Four people, twelve agents, vague tickets

   Two to four people each launching agents at once on one origin, from named tasks to vague tickets in a 654-file repo with decoys and a task that depends on a teammate's unpushed work.

   4–12 per tier, each tier with and without Cork agents · 7.8M vs 14.7M tokens for twelve agents, every task correct; 7.4M with the Ponytail plugin added

   47%

   fewer tokens on the hardest tier, same run](https://corkcli.com/benchmarks/study-3-teams-easy-to-hard)
4. [Study 4

   Oct 3, 2026

   Twelve agents, four rounds

   Study 3's hardest tier with one person's twelve agents at once, repeated four times with and without Cork, Cork's paid plan set up at the machine level.

   96 (4 rounds of 12, each way) agents · 13.1M vs 20.9M tokens; 47/48 tasks correct vs 46/48

   37%

   fewer tokens over four rounds](https://corkcli.com/benchmarks/study-4-twelve-agents-four-rounds)
5. [Study 5

   Oct 1, 2026

   Twelve agents with Cork's paid plan

   Study 3's hardest tier with twelve agents at once, with Cork's paid plan (prompts matched by meaning) and without.

   24 (12 + 12) agents · 5.5M vs 11.8M tokens; 152 vs 312 turns; 284s vs 373s

   53%

   fewer tokens, same tasks and model](https://corkcli.com/benchmarks/study-5-twelve-agents-53-percent)
6. [Study 6

   Oct 3, 2026

   Four people, twelve agents, four rounds

   Study 3's hardest tier with four people each on their own machine and sign-in, three agents each, repeated four times with and without Cork's paid plan.

   96 (4 rounds of 12, each way) agents · 13.8M vs 19.3M tokens; 47/48 tasks correct vs 47/48

   29%

   fewer tokens over four rounds](https://corkcli.com/benchmarks/study-6-four-people-twelve-agents)
7. [Study 7

   Oct 15–16, 2026

   From one agent to sixteen

   1, 2, 4, 8 and 16 agents at once on Study 3's hardest repo, five rounds at each size with and without Cork, to see whether the saving grows with the team.

   1–16 at once, 5 rounds at each size, each way agents · 24% at 1 agent to 47% at 16, almost all from landing; 151/155 tasks correct vs 151/155

   47%

   fewer tokens at 16 agents, five rounds](https://corkcli.com/benchmarks/study-7-scaling-to-sixteen-agents)
8. [Study 8

   Oct 16, 2026

   Sixteen agents: a regression, then the fixes

   Sixteen agents at once on a new, harder 527-file repo, three rounds without Cork, then three each on an unreleased build that regressed, the unreleased fix build, and unreleased v76 and v78 on the staging server.

   240 (3 rounds of 16, five ways) agents · first build 2% more tokens (37.6M vs 36.8M per round); fix build 29% fewer on a local server; v76 20% fewer on staging with three regressions; v78 fixed them: 26.1M, 46/48 tasks correct vs 48/48

   28.5%

   fewer tokens, latest run: v78 on the staging server](https://corkcli.com/benchmarks/study-8-sixteen-agents-a-regression)

## Questions, answered from the data

How do I reduce token usage when running several AI coding agents on one codebase?
:   Stop each agent from rediscovering the codebase. Cork gives every agent the exact signatures, callers and recent changes of the code its prompt names before it starts, and lets later agents replay work an earlier agent already verified. In a measured run of ten agents each way, agents on Cork used 5.3× fewer tokens per agent (240K vs 1,262K), with every task correct.

Does adding more AI agents make each one cheaper?
:   With shared context, yes: the first agent pays for discovery and the rest inherit it. In Cork's Study 2 the running cost per agent fell from 545K to 240K as ten agents joined, while agents without shared context stayed around 1,262K each. The total still grows with every agent; what falls is the cost of each one.

How were these token counts measured?
:   From the provider-reported usage in each agent's own session: input, cache writes, cache reads and output, once per assistant message. For all 20 agents in Study 2, summing the transcript gives exactly the total Claude Code reported. Correctness is the repository's own check, not the agent's summary.

Can two AI agents take the same migration number at the same time?
:   Not with Cork's claim tool: numbers are handed out by one atomic database statement. 500 concurrent claims from 50 sessions came back as 500 distinct numbers with no gaps in 1.33 seconds, and a file that takes someone else's claimed number is flagged as a collision.

What is Cork?
:   Cork is a shared context layer for teams whose code is edited by several AI coding agents and people at once. It connects over MCP and hooks to Claude Code, Codex, Cursor and other agents, shows every agent what the others are changing down to the symbol, catches collisions before they land, and remembers what earlier agents learned.

Is Cork cheaper than letting each agent search the repo with grep?
:   In tokens, yes: 5.3× fewer per agent in the latest same-run study. In dollars the gap is smaller because most of a long session is billed at the cache-read rate: $1.69 vs $4.91 for ten agents each way.

This page in HTML: https://corkcli.com/benchmarks