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, bash recovery, tool hallucination), estimated from millions of live Agent Mode sessions. Percent with 95% CI; not an Elo.

All results

50 rows

#	Model	Effort	Agent Arena
1	Opus 5.5	max	20.0% ±2.1
2	Gemini 4 Carbon	high	16.6% ±0.8
3	Sonnet 5.5	max	14.6% ±2.8
4	GPT-6 Astra	max	13.7% ±2.0
5	Fable 5.1	max	13.1% ±1.9
6	GPT-6.1 Sol	max	12.2% ±1.1
7	Gemini 4 Argon	high	10.6% ±0.8
8	GPT-6.1 Terra	max	10.2% ±1.2
9	Fable 5	high	7.8% ±1.1
10	Opus 5	max	7.6% ±1.2
11	GPT-6 Sol	max	6.9% ±1.6
12	Kimi K3.1	max	4.7% ±1.3
13	GPT-5.6 Sol	xhigh	4.6% ±1.0
14	Sonnet 5	high	2.8% ±1.2
15	Grok 4.7	xhigh	2.6% ±1.5
16	Gemini 3.8 Flash	high	2.6% ±0.8
17	Muse Spark 1.3	max	2.4% ±0.8
18	MiMo V2.6 Pro	high	2.4% ±0.9
19	DeepSeek V4.1 Flash	max	2.1% ±0.6
20	GPT-6 Luna	max	1.9% ±0.4
21	GLM 5.3	max	1.2% ±0.8
22	Hy4	high	0.9% ±0.7
23	Kimi K3	max	0.9% ±0.6
24	GLM 5.2	max	0.7% ±0.8
25	Haiku 5.5	high	0.7% ±1.0
26	MiMo V2.6 Flash	high	0.4% ±1.0
27	Qwen 3.8 Max	max	0.4% ±0.6
28	Grok 4.6	xhigh	0.3% ±1.1
29	Nemotron 3 Ultra 550B A55B	max	-15.6% ±3.2
30	DeepSeek V4 Pro	max	-0.3% ±1.7

Settings

Theme

Explore themes

Library

Bring your own theme

Mode

Agent Arena

Leaderboard

All results