# Harnesses Composite Index

> Coding agent harnesses compared on fixed test models and full model-plus-harness stacks.

- Canonical URL: https://prochoice.industries/harnesses
- Domain JSON endpoint: https://prochoice.industries/api/v1/domains/harnesses
- Epoch version: `harnesses-v1.0.0`
- Snapshot hash: `43237d0f0a7497fc`
- Updated: 2026-11-21
- Cost axis: Measured cost per run (Terminal-Bench 4.1 & VulcanBench SWE v4, USD, log scale)
- Evaluated entries: 25

## Headline picks

- Budget: OpenClaw at 38.7% and $0.39
- Best value: Pi at 44.8% and $0.52
- Top score: OpenCode at 48.0% and $0.57

## Pareto frontier

| Rank | Harness | Version | Vendor | Pass rate | Cost / run |
| :--- | :--- | :--- | :--- | :--- | :--- |
| 1 | OpenCode | v2.2.6 | Anomaly | 48.0% | $0.57 |
| 2 | Pi | v1.4.0 / v1.2.0 | Pi | 44.8% | $0.52 |
| 3 | Oh My Pi | v18.12.6 | Oh My Pi | 44.1% | $0.52 |
| 4 | OpenClaw | v2026.11.1 | OpenClaw | 38.7% | $0.39 |

## Controlled benchmark sources

- FrontierHarness Eval v1.1 (22 harnesses across 5 open control models): https://frontierharness.org/
- Harness Leaderboard (12 harnesses across 4 open control models): https://joelniklaus-harness-leaderboard.hf.space/

Full JSON dataset: https://prochoice.industries/api/v1/snapshot
