Four models on one prompt: who actually builds the pelican-on-a-bicycle machine
The same prompt was run on GPT-6 Astra, Claude Opus 5, Gemini 3.7 Flash and GLM 5.3 Flash to see which one's leg inverse kinematics breaks first.
the pelican-on-a-bicycle test is dumb on purpose. if the legs don’t inverse-kinematic, the model didn’t build a machine. it drew a picture. Lonely ran the same prompt on four models through @ZenMuxAI. one key. no four dashboards. GPT-6 Astra: crank and pedals lock. 12 tool calls, 9,051 output tokens, working loop first pass. Claude Opus 5: best color and taste, but cadence gets floaty at higher RPM. Gemini 3.7 Flash: constraints hold, controls feel snappy, it’s fast. GLM 5.3 Flash: fun comic look, but the leg linkage breaks under stress. that’s the real test. same prompt. different models. compare the failure.
This post quotes another post: x.com/lonely__mh/status/2097587056775934398