Which AI model should you use?
Compass measured 12 image and 5 video models on 96 real requests and sends you to the winner, with proof.
Nearest tested requests, made with this model
Compass could not tell what you want to make. Pick the closest task:
12 tasks, one method
Image models got the same 80 requests, video models the same 16; the best model per task is the judge's most likely first pick.
-
Product photo
GPT Image 2.5 Sunburst standard
21% chance of ranking first10 requestsjudge agreement 0.63
-
Logo with text
GPT Image 2.5 Sunburst standard
27% chance of ranking first10 requestsjudge agreement 0.60
-
Poster and flyer
GPT Image 2 premium
25% chance of ranking first10 requestsjudge agreement 0.67
-
Portrait and headshot
GPT Image 2.5 Sunburst standard
25% chance of ranking first10 requestsjudge agreement 0.59
-
Anime and illustration
GPT Image 2.5 Sunburst standard
16% chance of ranking first10 requestsjudge agreement 0.39
-
Interior and architecture
Hunyuan V3 standard
30% chance of ranking first10 requestsjudge agreement 0.50
-
Fantasy and concept art
Nano Banana 2 standard
29% chance of ranking first10 requestsjudge agreement 0.52
-
Stickers and icons
GPT Image 2 premium
30% chance of ranking first10 requestsjudge agreement 0.63
Video Video, early results
5 video models on 16 requests, 4 per task. See the all-video ranking, the stronger signal.
-
Product video
Seedance 2.5 premium
30% chance of ranking first4 requestsjudge agreement 0.70
-
Cinematic scene
Wan 2.7 standard
57% chance of ranking first4 requestsjudge agreement 0.85
-
Animated character
Veo 3.1 Fast fast
30% chance of ranking first4 requestsjudge agreement 0.75
-
People in motion
Wan 2.7 standard
28% chance of ranking first4 requestsjudge agreement 0.50
Same requests to every model of a kind (80 image, 16 video), a blind AI judge run twice, Plackett–Luce ratings with error bars.
Read the method and its limits