Benchmarks · agentic

Agent Arena · Shell recovery

Agent Arena signal: recovering in few steps after a failed shell command. A model's contribution to real agentic sessions, in percentage points.

All results

49 rows

#	Model	Effort	Score
1	Sonnet 5.5	max	15.0% ±1.7
2	Opus 5	max	13.6% ±0.8
3	Opus 5.5	high	13.6% ±0.8
4	Opus 5	high	12.6% ±0.8
5	Fable 5.1	max	11.6% ±2.3
6	Fable 5	high	9.6% ±1.2
7	GPT-5.5	xhigh	9.5% ±1.2
8	DeepSeek V4.1 Flash	max	7.9% ±0.3
9	Sonnet 5	high	7.4% ±1.2
10	Opus 4.8	high	7.2% ±1.9
11	Grok 4.7	xhigh	7.2% ±1.0
12	GPT-6 Sol	max	7.0% ±1.9
13	Hy4	high	6.5% ±0.5
14	Muse Spark 1.3	max	6.1% ±0.6
15	Muse Spark 1.2	xhigh	5.8% ±0.9
16	Grok 4.6	xhigh	5.7% ±1.0
17	GPT-6 Astra	max	5.5% ±1.7
18	Inkling Small	max	5.4% ±1.5
19	GPT-6 Luna	max	5.2% ±0.8
20	Grok 4.5	max	5.1% ±2.0
21	GPT-6.1 Sol	max	4.5% ±1.6
22	DeepSeek V4 Pro	high	4.4% ±1.3
23	Kimi K3	max	4.4% ±0.5
24	Qwen 3.8 Max	max	2.8% ±0.6
25	MiMo V2.6 Pro	high	2.3% ±2.0
26	Mistral Large 4	high	1.8% ±3.0
27	GPT-5.4	high	1.2% ±1.4
28	GPT-5.6 Terra	xhigh	0.9% ±1.8
29	GPT-5.6 Sol	xhigh	0.7% ±1.4
30	GLM 5.2	max	0.7% ±0.8