LMArena Agent Arena Bash Recovery
Agents · 2026-06-04
Historical signed percentage-point Agent Arena diagnostic measuring recovery after shell or command-line failures. Values can be negative relative to the leaderboard baseline. The publisher replaced bash_recovery_steps with the distinct Bash Reliability signal in October 2026; archived observations retain this original metric and publication date.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Sonnet 5.5 | 15.0 |
| Claude Opus 5.5 | 15.0 |
| Claude Opus 5 | 13.6 |
| Claude Fable 5.1 | 11.6 |
| Claude Fable 5 | 9.6 |
| GPT-5.5 | 9.5 |
| DeepSeek V4.1 Flash | 7.9 |
| Sonnet 5 | 7.4 |
| Opus 4.8 | 7.2 |
| Grok 4.7 | 7.2 |
| GPT-6 Sol | 7.0 |
| Hy4 Preview | 6.5 |
| Muse Spark 1.3 | 6.1 |
| Muse Spark 1.2 | 5.8 |
| Grok 4.6 | 5.7 |
| GPT-6 Astra | 5.5 |
| Inkling-Small | 5.4 |
| Grok 4.5 | 5.1 |
| GPT-6.1 Sol | 4.5 |
| DeepSeek V4 Pro 0813 | 4.4 |
| Kimi K3 | 4.4 |
| Qwen3.8-Max | 2.8 |
| MiMo-V2.6-Pro | 2.3 |
| Mistral Large 4 Preview | 1.8 |
| GPT-5.4 | 1.2 |
| GPT-5.6 Terra | 0.9 |
| GPT-5.6 Sol | 0.7 |
| GLM-5.2 | 0.7 |
| MiniMax M3 | 0.6 |
| GPT-5.6 Luna | 0.5 |
| Step 5 Preview | 0.2 |
| Inkling | 0.1 |
| GLM-5.3 | -0.1 |
| Gemini 3.8 Flash | -0.7 |
| MiMo-V2.6-Flash | -1.1 |
| Qwen3.7-Max (source unspecified) | -1.6 |
| Muse Spark 1.1 | -2.3 |
| Qwen3.8-Flash-Next | -2.6 |
| Gemini 3.7 Flash | -5.4 |
| GLM-5.3 Flash | -5.7 |