Evidence-based evaluation (not tested by Loadout): Terminal-Bench 4.0
The public Terminal-Bench 4.0 leaderboard (35 entries, fetched 2026-10-19) lists only two of our seven ranked coding agents: Claude Code and Codex. Best Claude Code entry: 64.85% ± 3.11 (Opus 5.5, max effort, 2026-09-22). Best Codex entry: 58.18% ± 2.79 (GPT-6 Astra, max) and 58.18% ± 3.11 (GPT-6.1 Sol, max). The 95% intervals do not overlap (61.74 vs 61.29), so the evidence is consistent with the editorial order (#1 Claude Code, #2 Codex). Cursor, GitHub Copilot, Devin, Cline and OpenHands have no entries, so there is no benchmark evidence for them either way.
Rank impact: none. No score or rank was changed by this log.
Method
- Fetched the leaderboard page and decoded its embedded data (agent, model, reasoning effort, accuracy, 95% CI half-width, trials, date).
- Compared the best entry per ranked agent, noting the result mixes agent harness and model quality.
Results
| tool | best entry |
|---|---|
| Claude Code | Claude Code + Opus 5.5 (max): 64.85% ± 3.11, 330 trials, 2026-09-22 |
| OpenAI Codex | Codex + GPT-6 Astra (max): 58.18% ± 2.79; Codex + GPT-6.1 Sol (max): 58.18% ± 3.11 |
Not covered
- cursor
- github-copilot
- devin
- cline
- openhands (no Terminal-Bench 4.0 entries)
Limitations
Third-party results, not reproduced by Loadout. Scores reflect agent+model pairs, not the agent alone. One benchmark (terminal tasks) only. Used as supporting evidence for the existing order; no numeric Loadout Score is derived from it.