Benchmarks · agentic

Agent Arena · Praise vs complaint

Agent Arena signal: positive responses relative to negative ones. A model's contribution to real agentic sessions, in percentage points.

All results

46 rows

#	Model	Effort	Score
1	Opus 5.5	max	46.0% ±8.2
2	Gemini 4 Carbon	high	37.9% ±3.7
3	GPT-6.1 Sol	max	27.7% ±3.9
4	GPT-6 Astra	max	27.1% ±7.3
5	Fable 5.1	max	27.1% ±6.9
6	Gemini 4 Argon	high	24.0% ±2.8
7	Sonnet 5.5	max	23.8% ±10.3
8	Fable 5	high	15.5% ±3.5
9	GPT-6 Sol	max	13.4% ±5.6
10	Opus 5	max	11.0% ±4.1
11	GPT-5.6 Sol	xhigh	6.6% ±3.3
12	Gemini 3.8 Flash	high	5.1% ±2.5
13	GLM 5.2	max	3.8% ±2.3
14	GLM 5.3	max	3.7% ±2.3
15	MiMo V2.6 Pro	high	3.4% ±2.7
16	DeepSeek V4.1 Flash	max	2.4% ±1.1
17	Grok 4.7	xhigh	2.3% ±4.6
18	Muse Spark 1.3	max	2.2% ±2.5
19	Hy4	high	1.6% ±2.1
20	Kimi K3	max	1.5% ±1.7
21	Sonnet 5	high	1.0% ±3.6
22	DeepSeek V4 Pro	max	0.6% ±5.2
23	Qwen 3.8 Max	max	0.4% ±1.9
24	MiMo V2.6 Flash	high	0.3% ±3.6
25	GPT-6 Luna	max	0.1% ±1.3
26	Nemotron 3 Ultra 550B A55B	max	-21.0% ±7.0
27	Haiku 5.5	high	-3.0% ±3.3
28	Gemini 3.1 Pro	high	-2.4% ±3.4
29	Gemini 3.7 Flash	high	-1.6% ±2.4
30	GLM 5.3 Flash	max	-4.3% ±1.9