What a coding harness actually costs you

Everyone compares models. Almost nobody compares harnesses — yet the harness decides how many tokens your task burns, and tokens are what you pay for.

On our fixed benchmark suite, two popular harnesses produced this:

Harness	Task score	Tokens used
3code	9/10	5.1M
Pi	6/10	4.6M

Same models, same tasks. One harness spent 10% fewer tokens and scored 50% lower. That gap — score per million tokens — is the metric this site exists to measure, and no leaderboard tracks it.

Why it matters

At DeepSeek prices, 5M tokens is pennies. At frontier-model prices it's real money, and it compounds daily for anyone running agents all day. The harness choice changes your bill more than the model choice — but model choice gets all the articles.

Methodology

Fixed task suite, fixed models, fresh session per task, full token accounting including cache reads. Raw logs published with every run. Full methodology post coming next.