Benchmarks › Terminal-Bench
Terminal-Bench
By Stanford University, Laude Institute and the open-source community ↗Results by Artificial Analysis ↗
Agents complete tasks in a terminal
63.6top score · Claude Sonnet 5.5
132models ranked
79open-weights models
Percentage of tasks solvedhigher is better
Oct 20, 2026last updated
How the test works
1
Task
66 tasks
- What the model doesAgents complete tasks in a terminal
- Kinds of tasksTerminal task execution
- In → outText → Terminal actions
2
Run
3 runs per task
- Agent / harnessmini-swe-agent
- Step limit500 steps
- Token budgetNon-reasoning: up to 16,384 output tokens, adjusted to model limits; reasoning: model-specific maximum
3
Score
Percentage of tasks solved
- Graded byAll task tests must pass; pass/fail verification
- Better isHigher
Sources: artificialanalysis.ai
Terminal-Bench 4.0 leaderboardtop 10 of 132
Full leaderboard with filters →| Model | Terminal-Bench 4.0 | Per task | Price | Speed | Supports | Context |
|---|---|---|---|---|---|---|
63.6max | $5.46 | $2 / $10 | 137 | 1M | ||
59.6max | $5.98 | $4 / $20 | 96 | 1M | ||
59.6xhigh | $2.31 | $10 / $50 | 61 | 1.05M | ||
57.1high | $1.99 | $2 / $10 | 84 | 1M | ||
57.1 | — | — | 1M | |||
56.1max | $0.72 | $2 / $10 | 52 | 1.05M | ||
55.1xhigh | $5.98 | $10 / $50 | 64 | 1M | ||
52max | $4.52 | $4 / $20 | 161 | 1M | ||
49max | $5.86 | $5 / $25 | 55 | 1M | ||
43.9max | $1.04 | $2 / $10 | 128 | 1.05M |
Versions and boards on Model Menu
| Board | Models |
|---|---|
Terminal-Bench 4.0Shown above | 132 |
| 242 | |
| 33 |
Terminal-Bench questions
What is Terminal-Bench?›
Agents complete tasks in a terminal. It has 66 tasks. Created by Stanford University, Laude Institute and the open-source community.
Which model is best on Terminal-Bench?›
Claude Sonnet 5.5 leads Terminal-Bench 4.0 with 63.6, ahead of Claude Opus 5.5 (59.6) and GPT-6 Astra (59.6).
What is the best open-weights model on Terminal-Bench?›
GLM 5.3 with 41.9, then MiMo v2.6 Pro (34.8).
How is Terminal-Bench scored?›
Percentage of tasks solved; all task tests must pass; pass/fail verification; 3 runs per task per task.
Which versions of Terminal-Bench are there?›
Terminal-Bench 4.0 (132 models), Terminal-Bench Hard (242 models), Terminal-Bench Science 0.1 (33 models).