Inference
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.43
Top resultBest result on this test.13.1Claude Opus 5
Top-3 spreadPoints from first to third.3.4
UpdatedDate the source changed the results.15 Sept 2026

Result and price

-30-20-1001020$0.3$1$3$10$30$100RESULTOUTPUT PRICE PER 1M TOKENS · LOG SCALE

Results

43 results
#ModelResultThe score from the source.
01Claude Opus 5max effort13.1
02Claude Fable 5.1max effort12.6
03GPT-5.5xhigh effort9.6
04Claude Fable 5high effort9.1
05DeepSeek V4.1 Flashmax effort7.6
06Claude Opus 4.8high effort7.5
07Hy4 preview7.4
08Claude Sonnet 5high effort7.4
09Muse Spark 1.2xhigh effort7.3
10GPT-6 Astramax effort7.3
11Inkling-Small6.9
12Grok 4.6xhigh effort6.6
13Grok 4.56.6
14DeepSeek V4 Pro (High) (0813)6.2
15Muse Spark 1.3max effort5.9
16Claude Sonnet 4.65.0
17GPT-5.4high effort4.6
18GPT-5.6 Solxhigh effort4.3
19Qwen3.8max effort4.1
20GPT-5.6 Lunaxhigh effort3.7
21DeepSeek V4 Pro 04233.2
22GPT-5.6 Terraxhigh effort3.0
23Inkling2.0
24GLM-5.2max effort1.9
25Gemini 3.8 Flashhigh effort1.9
26MiniMax M31.6
27Deepseek V4 Flash (High) (20260731)1.5
28Muse Spark 1.1-0.6
29Qwen3.7max effort-0.8
30GLM-5.3max effort-1.0
31Qwen3.8-Flash-Next-2.0
32Gemini 3.7 Flashhigh effort-2.1
33Hy3-3.9
34Qwen3.7 Plus-4.4
35GLM-5.3-Flash-4.6
36MiMo-V2.5-Pro-4.9
37Qwen3.8-27B-5.2
38Gemini 3.6 Flashhigh effort-12.4
39Mistral Medium 3.5-15.8
40Gemini 3.1 Pro Preview-20.4
41Solar Pro 4-24.2
42MiniMax M2.7-26.3
43Gemini 3.5 Flash-Lite-28.7