Benchmarks · agentic

Agent Arena · Steerability

Agent Arena signal: how well it lands user corrections. A model's contribution to real agentic sessions, in percentage points.

All results

46 rows

#	Model	Effort	Score
1	Gemini 4 Carbon	high	18.4% ±1.0
2	Opus 5.5	max	18.0% ±1.9
3	Fable 5.1	max	14.9% ±2.1
4	Gemini 4 Argon	high	13.7% ±0.9
5	GPT-6 Astra	max	13.6% ±2.1
6	GPT-6.1 Sol	max	13.5% ±1.1
7	Sonnet 5.5	max	10.7% ±3.0
8	GPT-6 Sol	max	10.6% ±1.8
9	Fable 5	high	9.3% ±1.4
10	GPT-5.6 Sol	xhigh	6.4% ±1.4
11	Opus 5	max	5.7% ±1.5
12	Sonnet 5	high	4.8% ±1.6
13	GPT-6 Luna	max	4.5% ±0.5
14	Muse Spark 1.3	max	4.1% ±1.1
15	Haiku 5.5	high	4.1% ±1.3
16	MiMo V2.6 Flash	high	2.6% ±1.5
17	MiMo V2.6 Pro	high	1.2% ±1.2
18	GPT-5.6 Terra	xhigh	1.1% ±1.1
19	Hy4	high	0.8% ±0.9
20	Kimi K3	max	0.4% ±0.8
21	Grok 4.6	xhigh	0.4% ±1.6
22	GLM 5.2	max	0.4% ±1.2
23	Gemini 3.8 Flash	high	0.4% ±1.1
24	DeepSeek V4.1 Flash	max	0.1% ±0.5
25	Qwen 3.8 Max	max	0.1% ±0.9
26	Nemotron 3 Ultra 550B A55B	max	-18.8% ±6.6
27	DeepSeek V4 Pro	max	-0.6% ±2.4
28	Gemini 3.1 Pro	high	-6.9% ±2.7
29	Gemini 3.7 Flash	high	-5.1% ±1.4
30	GLM 5.3 Flash	max	-3.8% ±1.1