Inference
EN
LMArena

Agent Arena command recovery

Estimated effect of the model on how quickly an agent recovers from a failed shell command.

ModelsModels with a result.50
Top resultBest result on this benchmark.15.0Claude Sonnet 5.5
Top-3 spreadPoints from first to third.1.5
UpdatedDate the source changed the results.8 Oct 2026

Result and price

47 of 50 results

Results

50 results
#ModelResultThe score from the source.
01Claude Sonnet 5.5max effort15.0
02Claude Opus 5max effort13.6
02Claude Opus 5.5high effort13.6
04Claude Fable 5.1max effort11.6
05Claude Fable 5high effort9.6
06GPT-5.5xhigh effort9.5
07DeepSeek V4.1 Flash 1017max effort7.9
08Claude Sonnet 5high effort7.4
09Claude Opus 4.8high effort7.2
09Grok 4.7xhigh effort7.2
11GPT-6 Solmax effort7.0
12Hy4 preview6.5
13Muse Spark 1.3max effort6.1
14Muse Spark 1.2xhigh effort5.8
15Grok 4.6xhigh effort5.7
16GPT-6 Astramax effort5.5
17Inkling-Small5.4
18GPT-6 Lunamax effort5.2
19Grok 4.55.1
20GPT-6.1 Solmax effort4.5
21DeepSeek V4 Pro (High) (0813)4.4
21Kimi K3max effort4.4
23Qwen3.8 Max2.8
24MiMo-V2.6-Pro2.3
25Mistral Large 41.8
26GPT-5.4high effort1.2
27GPT-5.6 Terraxhigh effort0.9
28GPT-5.6 Solxhigh effort0.7
28GLM-5.2max effort0.7
30MiniMax M30.6
31GPT-5.6 Lunaxhigh effort0.5
32Step 5 Preview0.2
33Inkling0.1
34GLM-5.3max effort−0.1
35Gemini 3.8 Flashhigh effort−0.7
36MiMo-V2.6-Flash−1.1
37Qwen3.7 Max−1.6
38Muse Spark 1.1−2.3
39Qwen3.8-Flash-Next−2.6
40Gemini 3.7 Flashhigh effort−5.4
41GLM-5.3-Flash−5.7
42Qwen3.8-27B−5.9
43Qwen3.7 Plus−6.3
44Hy3−7.2
45MiMo-V2.5-Pro−7.5
46Gemini 3.6 Flashhigh effort−18.2
47Mistral Medium 3.5−20.4
48Gemini 4 Argonhigh effort−21.4
49Gemini 3.1 Pro Preview−23.3
50Solar Pro 4−25.6