AI coding agents still need human cleanup on nearly three out of four tasks, according to a new benchmark from researchers at Alibaba and several Asian universities.
The team built Co-Design Bench from real user change requests. Agents must rework existing web apps based on those requests, sometimes using images, audio recordings, videos, or documents as input. The researchers favored cases where users had to rephrase requests or roll back changes, so the tasks range from small tweaks, such as pulling an amount from a receipt and updating dependent totals, to full project overhauls.
According to the paper, GPT-6 Astra was the best of the ten models tested. It meets about 70 percent of a task's requirements on average, but it is fully correct in only 26.4 percent of tasks. In most cases, the agent does the bulk of the work but leaves at least one detail unfinished.
Grading is strict. One AI judge tests the finished app in a browser. A second judge rechecks every positive verdict and looks for mock-ups or hardcoded results. Points are awarded only for features that work in the app.
Claude Opus 5 and Qwen3.8-Max rank next, and Grok 4.6 places last. On larger overhauls, the much smaller Qwen3.