Vals AI Terminal-Bench 2.0: leaderboard
89 real-world terminal tasks in sandboxed environments including model training, system administration, and file operations across easy-to-hard difficulty levels.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around January 2027. 67 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 73.2 | #31 |
| 2 | Claude Opus 4.8 | 70.04 | #37 |
| 3 | Claude Opus 4.7 | 68.54 | #49 |
| 4 | Gemini 3.1 Pro (Preview) | 67.42 | #58 |
| 5 | Gemini 3.5 Flash | 67.42 | #60 |
| 6 | GPT-5.3 Codex | 64.05 | #69 |
| 7 | Claude Sonnet 4.6 | 59.55 | #90 |
| 8 | Muse Spark | 59.55 | #64 |
| 9 | Qwen 3.7 Max | 59.18 | #75 |
| 10 | Claude Opus 4.6 (Thinking) | 58.43 | #63 (Claude Opus 4.6) |
| 11 | GPT-5.4 (2026-03-05) | 58.43 | #59 |
| 12 | Claude Opus 4.5 (20251101) | 58.43 | #74 |
| 13 | Kimi K2.6 | 57.3 | #104 |
| 14 | DeepSeek V4 Pro | 56.18 | #102 |
| 15 | Gemini 3 Pro (Preview) | 55.06 | #66 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-terminal-bench-2-0 · How It Works · Data refreshed daily, snapshot 2026-10-20.