LMArena Agent Arena Work Bash Reliability
Agents · 2026-08-14
Signed percentage-point Agent Arena Work category diagnostic signal for successful bash commands, excluding failures outside the model's control, recomputed on sessions tagged as office work such as document creation, professional research, planning and professional writing. Atlas keeps the publisher's raw bash_reliability_no_prereq sign, so values are baseline-relative, not absolute success probabilities, and may be negative relative to the category baseline; reported confidence intervals are not standard errors.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 15.5 |
| Sonnet 5 | 15.5 |
| GPT-5.5 | 12.5 |
| Claude Sonnet 5.5 | 12.5 |
| GPT-5.6 Sol | 11.7 |
| Claude Fable 5 | 11.2 |
| Claude Opus 5 | 10.9 |
| GPT-6 Sol | 10.6 |
| GPT-6 Luna | 10.4 |
| Gemini 3.8 Flash | 10.2 |
| Claude Fable 5.1 | 9.9 |
| Grok 4.6 | 9.6 |
| GPT-5.4 | 9.5 |
| Grok 4.7 | 9.0 |
| GPT-6 Astra | 7.9 |
| Muse Wonder Full | 7.5 |
| Claude Haiku 5.5 | 7.3 |
| GPT-5.6 Luna | 6.6 |
| Gemini 3.7 Flash | 6.1 |
| Muse Wonder | 6.0 |
| DeepSeek V4.1 Flash 1017 | 5.2 |
| Qwen3.7-Max (source unspecified) | 4.3 |
| DeepSeek V4 Pro 0813 | 3.9 |
| MiMo-V2.6-Pro | 3.6 |
| Inkling-Small | 3.1 |
| Qwen3.7-Plus (source unspecified) | 2.7 |
| Inkling | 2.2 |
| GPT-6.1 Sol | 2.1 |
| Muse Spark 1.3 | 2.0 |
| Gemini 3.1 Pro Preview | 1.9 |
| GPT-5.6 Terra | 1.4 |
| Mistral Large 4 Preview | 0.8 |
| Qwen3.8-Max | 0.3 |
| MiniMax M3 | 0.1 |
| DeepSeek V4.1 Flash | 0.0 |
| MiMo-V2.5-Pro | -0.4 |
| Hy4 Preview | -1.6 |
| MiMo-V2.6-Flash | -2.0 |
| Muse Spark 1.1 | -2.3 |
| Kimi K3 | -2.7 |