FrontierHarness Eval
FrontierHarness Eval is an open benchmark of AI coding agent harnesses. It runs 21 harness configurations, including Claude Code, Codex CLI and OpenCode, on 5 models and 62 tasks in software engineering, science, data, systems, ML and security, and compares their pass rate, cost and speed.
Every harness runs the same 62 pinned tasks with each model, 4 scored attempts per task. Only verifier passes count; model failures and timeouts score 0. Infrastructure failures are excluded from scores and tracked separately.
The leaderboard and its charts need JavaScript.
FrontierHarness Eval