Terminal-Bench 2.0 - Hard
Agents · 2026-06-04
The Hard split of Terminal-Bench 2.0. This child benchmark separates a source-reported subtask or subtrack from the parent aggregate so scores at different grains do not share one benchmark_slug.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 | 60.5 |
| Opus 4.7 | 56.7 |
| Opus 4.8 | 54.4 |
| Gemini 3.1 Pro Preview | 53.3 |
| Gemini 3.5 Flash | 53.3 |
| GPT-5.2 | 50.0 |
| GPT-5.3-Codex | 50.0 |
| Gemini 3 Flash Preview | 43.3 |
| Sonnet 4.6 | 40.0 |
| Muse Spark | 40.0 |
| Qwen3.7-Max | 40.0 |
| Gemini 3 Pro Preview | 36.7 |
| GLM-5.1 | 36.7 |
| Kimi K2.6 | 36.7 |
| Opus 4.5 | 33.3 |
| DeepSeek-V4-Pro | 33.3 |
| GPT-5.4 | 33.3 |
| Opus 4.6 | 30.0 |
| GPT-5.1 | 26.7 |
| MiniMax M2.7 | 26.7 |
| Qwen3.6 Plus (2026-04-02) | 26.7 |
| GPT-5.4 Mini | 24.4 |
| Grok 4.3 | 24.4 |
| Sonnet 4.5 | 23.3 |
| Gemma 4 31B IT | 23.3 |
| GLM-4.7 | 23.3 |
| GLM-5 | 23.3 |
| GPT-5.4 Nano | 21.7 |
| Kimi K2.5 | 20.0 |
| Kimi K2 Instruct | 20.0 |
| MiniMax M3 | 20.0 |
| Qwen3.6 27B | 20.0 |
| Qwen3.6-Max-Preview | 20.0 |
| GPT-5 | 16.7 |
| GPT-5 Mini | 16.7 |
| Grok 4.20 | 16.7 |
| Kimi K2 Thinking | 16.7 |
| MiniMax M2.5 | 16.7 |
| Qwen3.5 Plus (2026-02-15) | 16.7 |
| Haiku 4.5 | 13.3 |