Benchmarks · agentic

Agent Arena

Arena.ai Agent Arena: causal 'net improvement' a model contributes to real-world agentic sessions (task success, praise vs complaint, steerability, shell commands, tool hallucination), estimated from millions of live Agent Mode sessions. Percent with 95% CI; not an Elo.

All results

51 rows

#	Model	Effort	Agent Arena
1	Opus 5.5	max	20.1% ±2.1
2	Gemini 4 Carbon	high	16.7% ±0.8
3	Sonnet 5.5	max	15.4% ±2.8
4	GPT-6 Astra	max	14.7% ±2.1
5	Fable 5.1	max	13.9% ±1.9
6	GPT-6.1 Sol	max	12.5% ±1.1
7	GPT-6.1 Terra	max	11.4% ±1.4
8	Gemini 4 Argon	high	10.5% ±0.8
9	Fable 5	high	7.8% ±1.1
10	Opus 5	max	7.7% ±1.2
11	GPT-6 Sol	max	6.8% ±1.5
12	Hy4	high	5.4% ±1.7
13	Kimi K3.1	max	5.3% ±1.3
14	GPT-5.6 Sol	xhigh	4.6% ±1.0
15	MiMo V2.6 Pro	max	2.7% ±0.9
16	Grok 4.7	xhigh	2.7% ±1.5
17	Gemini 3.8 Flash	high	2.6% ±0.8
18	Muse Spark 1.3	max	2.5% ±0.8
19	Sonnet 5	high	2.4% ±1.2
20	DeepSeek V4.1 Flash	max	2.4% ±0.6
21	GPT-6 Luna	max	1.9% ±0.4
22	GLM 5.3	max	1.1% ±0.8
23	Kimi K3	max	0.8% ±0.6
24	Hy4 preview	high	0.7% ±0.7
25	GLM 5.2	max	0.6% ±0.8
26	Haiku 5.5	high	0.6% ±1.0
27	Qwen 3.8 Max	max	0.5% ±0.7
28	Grok 4.6	xhigh	0.5% ±1.2
29	MiMo V2.6 Flash	high	0.2% ±1.1
30	Nemotron 3 Ultra 550B A55B	max	-15.6% ±3.3

Settings

Theme

Explore themes

Library

Bring your own theme

Mode

Agent Arena

Leaderboard

All results