AI Coding Harness Leaderboards | Claude Code vs Codex vs Terminus 2 vs Cursor CLI

AI Coding Harness Leaderboards

loading snapshot...

cross-harness, cross-model agentic-coding scores

Same model, different harnesses, very different results. The harness owns context curation, tool design, retry policy, and verifier integration, so the same model can sit several points apart depending on which scaffold calls it, and the gap column below is the live measure of that. This page tracks the harness gap as a first-class number. Every row is an upstream Terminal-Bench row with its 95% confidence interval. We do not re-run benchmarks, and no vendor headline figure is promoted into the score column. Note that the two revisions are not comparable: 4.0 is the current and much harder board, 2.1 is the older and largely saturated one.

The harness is the agent scaffold that wraps a base language model: the system prompt, the tools, the file-edit format, the retry and verifier loop, the context-management strategy, and the way errors are surfaced back to the model.

Claude Code, Anthropic's CLI-native agent, default Opus or Sonnet
Codex, OpenAI's terminal agent, default GPT-5.x family
Terminus 2, the Terminal-Bench reference scaffold, model-agnostic
Cursor CLI, the command-line agent from the Cursor editor
Gemini CLI, Google's terminal agent, default Gemini family
Grok Build, xAI's coding agent
mini-SWE-agent, the minimal bash-only scaffold from the SWE-bench team