Co-Design Bench: best coding agent fully finishes only 26.4% of real user change requests
Alibaba and university researchers built Co-Design Bench from real change requests on existing web apps (some with image, audio, video, or document inputs), and top model GPT-6 Astra met about 70% of each task's requirements but got only 26.4% fully right. The same model's score also swung with the harness, with Qwen3.8-27B scoring 75% or 56% on large overhauls depending on the environment.