Terminal-Bench 4.0
Agents · 2026-08-28
Terminal-Bench 4.0 measures agents completing terminal tasks in Harbor. This major revision removes eight tasks, fixes nineteen tasks, calibrates CPU and memory resources, and gives each task an eight-hour agent timeout. Scores are mean resolution rates across repeated trials; they are not directly comparable to earlier Terminal-Bench task sets.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 58.2 |
| Claude Fable 5.1 | 57.9 |
| Claude Opus 5 | 53.9 |
| Claude Fable 5 | 44.5 |
| GLM-5.3 | 41.8 |
| Grok 4.7 | 37.6 |
| GPT-5.6 Sol | 37.3 |
| Muse Spark 1.3 | 27.8 |
| GPT-5.6 Terra | 26.3 |
| MiMo-V2.6-Pro | 24.7 |
| Opus 4.8 | 23.6 |
| MiMo-V2.6-Flash | 21.2 |
| Grok 4.6 | 20.3 |
| GLM-5.3 Flash | 19.7 |
| Gemini 3.8 Flash | 19.1 |
| GPT-5.6 Luna | 17.3 |
| Kimi K3 | 12.6 |
| Grok 4.5 | 12.4 |
| Sonnet 5 | 12.4 |
| DeepSeek V4.1 Flash | 11.6 |
| Gemini 3.7 Flash | 11.2 |
| DeepSeek-V4-Flash-0731 | 9.1 |
| Muse Spark 1.2 | 5.6 |
| Hy4 Preview | 5.1 |
| Gemini 3.6 Flash | 4.5 |
| Gemini 3.5 Flash | 4.0 |
| Qwen3.8 27B | 4.0 |
| Inkling-Small | 1.5 |
| DeepSeek V4 Pro 0813 | 1.0 |
| Inkling | 0.0 |
| Mercury 2.5 | 0.0 |