Agentic benchmark
Terminal-Bench 2.1 (Vals) leaderboard
Terminal-Bench 2.1, Vals AI run. Every model the catalog carries a published Terminal-Bench 2.1 (Vals) value for, ranked by that value.
Vals AI’s independent run of the Terminal-Bench 2.1 terminal-task suite with published easy, medium, and hard splits.
Terminal-Bench 2.1 (Vals) ranking
53 models with a published Terminal-Bench 2.1 (Vals) value, ordered by that value, highest first. Models the source has not scored on this benchmark are not listed — they are unmeasured, not last.
| Rank | Model | Provider | Task success rate |
|---|---|---|---|
| 1 |  GPT-6 Astra | OpenAI | 87.3 |
| 2 | GPT-5.6 Sol | OpenAI | 85.8 |
| 3 |  Claude Fable 5.1 | Anthropic | 85.0 |
| 4 | Claude Opus 5 | Anthropic | 84.6 |
| 5 |  Gemini 3.8 Flash | 81.3 | |
| 6 |  Kimi K3 | Moonshot AI | 80.9 |
| 7 | Claude Fable 5 | Anthropic | 80.5 |
| 8 | GPT-5.6 Luna | OpenAI | 79.0 |
| 9 |  Grok 4.6 | xAI | 78.3 |
| 10 | Gemini 3.7 Flash | 77.5 | |
| 10 | GPT-5.6 Terra | OpenAI | 77.5 |
| 12 | GPT-5.5 | OpenAI | 76.4 |
| 13 | Claude Sonnet 5 | Anthropic | 74.5 |
| 14 | Gemini 3.5 Flash | 74.2 | |
| 15 | Gemini 3.6 Flash | 73.8 | |
| 16 | Claude Opus 4.8 | Anthropic | 71.9 |
| 17 |  GLM-5.3 | Z.AI | 71.5 |
| 18 | Gemini 3.1 Pro | 70.8 | |
| 19 |  Muse Spark 1.2 | Meta | 69.7 |
| 20 | Muse Spark 1.1 | Meta | 69.3 |
| 21 | Claude Opus 4.7 | Anthropic | 68.5 |
| 22 | GLM-5.2 | Z.AI | 67.8 |
| 22 | Grok 4.5 | xAI | 67.8 |
| 24 |  Qwen3.8 Max | Alibaba | 67.4 |
| 25 |  DeepSeek V4 Flash 0731 | DeepSeek | 67.0 |
| 26 | GLM-5.3-Flash | Z.AI | 62.9 |
| 27 | Qwen3.7 Max | Alibaba | 61.0 |
| 28 |  MiMo-V2.5 | Xiaomi | 60.7 |
| 29 | Qwen3.8-27B | Alibaba | 58.4 |
| 30 | Claude Sonnet 4.6 | Anthropic | 57.3 |
| 30 | MiMo-V2.5-Pro | Xiaomi | 57.3 |
| 32 | GLM-5.1 | Z.AI | 56.9 |
| 33 | TMInkling-Small | Thinking Machines Lab | 55.1 |
| 34 | DeepSeek V4 Pro 0813 | DeepSeek | 54.7 |
| 34 | GPT-5.4 mini | OpenAI | 54.7 |
| 36 | Gemini 3 Flash | 53.9 | |
| 37 | Kimi K2.6 | Moonshot AI | 53.6 |
| 37 |  MiniMax M3 | MiniMax | 53.6 |
| 39 | Qwen3.6 Plus | Alibaba | 53.2 |
| 40 | Qwen3.7 Plus | Alibaba | 52.8 |
| 41 |  Nemotron 3 Ultra | NVIDIA | 50.9 |
| 42 | Gemini 3.5 Flash-Lite | 50.2 | |
| 42 |  Ling 3.0 Flash | InclusionAI | 50.2 |
| 44 | MiniMax M2.7 | MiniMax | 48.7 |
| 45 | TMInkling | Thinking Machines Lab | 47.6 |
| 46 | Grok 4.20 | xAI | 44.2 |
| 47 | Claude Haiku 4.5 | Anthropic | 43.8 |
| 48 | Grok 4.3 | xAI | 41.9 |
| 49 | GPT-5.4 nano | OpenAI | 41.6 |
| 50 |  Mistral Medium 3.5 128B | Mistral | 39.0 |
| 51 | Gemini 3.1 Flash-Lite | 34.1 | |
| 51 |  Laguna M.1 | Poolside | 34.1 |
| 53 | Laguna XS.2 | Poolside | 25.8 |
Evidence key: Observed
Rows are ordered by the value Vals AI Terminal-Bench 2.1, Vals AI run leaderboard published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.