This benchmark is retired. Its publisher released a new version; the scores below were recorded on the earlier version and are not comparable with it. Current scores are on the Terminal Bench 2.0 leaderboard.
A benchmark for evaluating the agentic ability of models to complete practical tasks by using tools in a terminal environment.
Current recorded best scores, ordered by model release date.
Hover or tap to inspect. Keyboard: โ / โ to move, Home / End to jump, Esc to reset.
GPT-5.1 Codex: 56.3%, released Nov 14, 2025
| Evaluation setting | Score |
|---|---|
| Normal (With Tools) | 27.0% |
| Thinking (With Tools)Best | 50.0% |
| Evaluation setting | Score |
|---|---|
| Normal (With Tools) | 43.3% |
| Extended (With Tools)Best | 46.5% |
| Evaluation setting | Score |
|---|---|
| Normal (With Tools) | 40.2% |
| Thinking (With Tools, 32K Budget)Best | 41.8% |
| Evaluation setting | Score |
|---|---|
| Normal (With Tools)anthropic.com | 35.5% |
| Deep Thinking (With Tools)BestOfficial leaderboard | 41.3% |
| Evaluation setting | Score |
|---|---|
| Normal (With Tools) | 23.0% |
| Thinking (With Tools)Best | 37.7% |
| Evaluation setting | Score |
|---|---|
| Normal (With Tools)Best | 36.7% |
| Thinking (With Tools) | 28.0% |
Source: checked 2026-10-05, unless a row names its own source.
Harness, budget and task set can change the same modelโs score.
GPT-5.1 Codex leads with 56.3%; the type of its source is not recorded. The best independently measured result is Claude Sonnet 4โs 41.3% (tbench.ai).
Provider-reported
35.5% ยท Normal
Independent
41.3% ยท tbench.ai ยท Deep Thinking
Independent: the benchmarkโs own leaderboard or a third-party evaluator. Provider-reported: the model makerโs model card, system card or launch post. When sources report the same result, the board shows the official leaderboard first, then the provider, then third-party tests.
The dataset license covers the test itself; each score source sets its own terms.
Published by Stanford University & Laude Institute ยท Apache-2.0 ยท Dataset licenseโ
Unless a license is listed, each sourceโs own terms apply.
Published by Anthropic
Source: Anthropic