Arena Vision · Captioning

Human preference on prompts with image inputs (multimodal understanding); Bradley-Terry rating. Only the Captioning category.

All results

31 rows

#	Model	Effort	Score
1	Gemini 3.1 Pro	high	1282 ±48
2	Gemini 3 Pro	high	1277 ±37
3	GPT-5.2	high	1262 ±58
4	Gemini 2.5 Pro	high	1251 ±23
5	GPT-5.1	high	1248 ±55
6	Gemma 4 26B	xhigh	1239 ±71
7	Gemini 3 Flash	high	1234 ±39
8	Gemini 3.1 Flash-Lite	high	1225 ±64
9	Gemini 2.5 Flash	medium	1211 ±26
10	Qwen 3	high	1211 ±46
11	GPT-4.1	high	1210 ±32
12	Gemini 3 Flash	minimal	1209 ±43
13	Kimi K2.5	max	1198 ±48
14	GPT-5	high	1196 ±36
15	GPT-4.1 Mini	high	1193 ±34
16	o4 mini	high	1185 ±32
17	o3	high	1183 ±30
18	Gemini 2.5 Flash Lite	none	1182 ±33
19	GLM 5 V	high	1180 ±74
20	GPT-5 Mini	high	1175 ±43
21	Qwen 3.5 397B-A17B	max	1174 ±57
22	GPT-5.2	xhigh	1172 ±56
23	Grok 4	high	1170 ±37
24	Gemma 4 31B	max	1170 ±55
25	Gemini 2.0 Flash	medium	1150 ±60
26	Mistral Small 3.1	none	1133 ±39
27	Mistral Medium 3.1	none	1130 ±31
28	Gemma 3 27B IT	medium	1117 ±39
29	Qwen 3.5 27B	high	1095 ±82
30	Mistral Medium 3	none	1090 ±50