The best AI coding agent fully finishes only 26% of real tasks

Researchers from Alibaba and several Asian universities built Co-Design Bench from 273 real user requests to change existing web apps — the kind of revision a user asks for in a live coding session — and graded the result against 2,331 hidden acceptance criteria written by experts.

The best of ten models tested, OpenAI's GPT-6 Astra, met 69.9% of the requirements on average but fully completed only 26.4% of tasks; Claude Opus 5 finished 18.3% and xAI's Grok 4.6 came last at 6.6%. The paper is an unreviewed preprint, and its tasks were deliberately chosen from cases where users had to rephrase a request or roll a change back.