New benchmark finds the best coding agent fully completes only about one in four real change requests

Researchers from Alibaba and several Asian universities have released Co-Design Bench, a test for AI coding agents built from real user change requests. The agents have to rework existing web apps, and some requests include images, audio recordings, videos or documents. Ten models were tested. The best, GPT-6 Astra, met about 70% of a task's requirements on average but produced a fully correct result in only 26.4% of tasks. In most cases the agent did most of the work but left at least one detail unfinished, so users would usually need to finish the job by hand. The authors built the test because older benchmarks no longer separate models well. Agents that can run, check and fix their own code score above 99% on some of them, including near-perfect results on WebCompass-edit, which makes differences between models hard to measure. For teams using coding agents, the results show how far a high benchmark score can be from finished work on real requests.