鹈鹕骑车测试进 Blender：GPT-6 Sol 对 Opus 5.5

Lakshya @Its_lakshya_ai原推发布于 2026.09.27

在 X 上打开原推

Opus 5.5 vs GPT-6 Sol: the pelican test just entered Blender 💀 The classic “pelican riding a bicycle” test has been everywhere in 2D. So someone asked the more interesting question: What happens if you make the models build it in 3D from scratch? Both got the exact same art brief: An Australian pelican on a teal touring bike, riding along a boardwalk at sunset, with a mackerel in its beak, water dripping from the fish, wing tips on the handlebars, and sunlight passing through the pouch. No downloaded assets. No ready-made scene. Just an empty Blender file and Python. And the results are surprisingly different. GPT-6 Sol finished faster, but the closer you look, the more the scene starts to fall apart. The leg IK gets weird, the feathers rotate with the legs, and the environment feels noticeably empty. Opus 5.5 took longer, but the final scene feels much more deliberate. The atmosphere is stronger, the colours work better, the environment has more life, and even the tiny splashes from the fish make the shot feel more complete. It still has a small rigging issue around the legs, but it's easy to miss unless you're looking for it. And that's what I find interesting about these tests. We’re moving past “can an AI make a 3D scene?” Both can. The next benchmark is much harder: Can it make something that survives a close-up? Opus 5.5: $48.11 GPT-6 Sol: $34.55 The pelican test just got a lot more interesting. 😭