Atlas

Benchmarks

← All benchmarks

Arena Vision — Captioning (Style Controlled)

Multimodal

Style-controlled Vision Arena rating for image-captioning prompts. Higher ratings indicate stronger preference in blind pairwise user votes after adjusting for response style.

Top models (higher is better)

ModelScore
Gemini 3.1 Pro Preview1274
Gemini 3 Pro Preview1269
GPT-5.21268
Gemini 2.5 Pro1251
Gemma 4 26B A4B1247
GPT-5.11240
Gemini 3.1 Flash-Lite Preview1226
Gemini 3 Flash Preview1226
Gemini 2.5 Flash1214
GPT-4.1 (2025-04-14)1208
GPT-4.1 Mini (2025-04-14)1196
Kimi K2.51196
GPT-51191
Gemini 2.5 Flash-Lite Preview 06-171185
O4 Mini (2025-04-16)1182
O3 (2025-04-16)1181
Grok 41168
GPT-5 Mini1167
Gemma 4 31B1165
Gemini 2.0 Flash 0011149
Mistral Small 3.1 24B Instruct 25031135
Mistral Medium 3.11129
Gemma 3 27B IT1118
Mistral Medium 31092
Mistral Small 3.2 24B Instruct 25061055
Loading Atlas data…