The model gets all the attention. The harness around it — prompts, tools, context management, when to stop — usually decides whether a coding agent actually ships work.

A new Harness Leaderboard from Joel Niklaus makes that measurable: same open models, same provider, same reasoning effort, same tasks. Only the coding-agent harness changes.

What FreePi is

FreePi is a free AI coding agent in your terminal. Ads pay for inference. Under the hood it runs Mario Zechner's Pi harness — the same open agent loop the leaderboard just put under a microscope.

What the board measured

Joel ran 28,849 trials across nine harnesses on GLM-5.3-Flash, DeepSeek-V4.1-Flash, and Flint, on Terminal-Bench 2.1 and DeepSWE (~$4,746 of inference). Headline score is resolve rate; cost is measured outside the harness.

Where Pi landed

Averaged over the three models:

  • Pi — 49.4% resolve at $0.29 per solved task
  • mini-swe-agent — 51.4% at $0.53 (Pi is statistically tied, ~45% cheaper)
  • OpenCode — $3.35 per solved task (11.7× Pi)

Pi sits in the top cluster and is the cheap strong option. The best harness still depends on the model, but the cost gap is hard to ignore.

Source announcement: Joel's post.

Why FreePi ships on Pi

We didn't invent a new agent loop. We ship Pi with free models, a GitHub sign-in, and no card required:

npx free-pi-cli@latest

When an independent board says the harness is near the top and thrifty, that's the product story: best harness we know how to ship, batteries included.

Wrapping up

Harness choice moves quality and cost as much as model choice. Pi cleared that bar. FreePi puts that harness in your terminal for free.