A new coding benchmark built from difficult real user requests found that the strongest tested agent completed only 26.4% of tasks perfectly. GPT-6 Astra met about 70% of requirements on average, but usually left at least one detail for a person to fix.
Researchers from Alibaba and several Asian universities created Co-Design Bench after older tests became saturated, with some agents scoring above 99%. Its tasks ask agents to modify existing web applications, sometimes using receipts, tables, audio, video or documents. The team favored cases where users had already needed to rephrase requests or undo failed changes.
Grading combines browser testing with a second judge that rechecks positive results and looks for hardcoded or cosmetic shortcuts. Performance fell as tasks mixed layout, design, interaction and content changes. Astra fully solved 41% of multimodal tasks but just under 18% of repository tasks, showing how an overall score can hide large differences between work types.
The software around the model also mattered. The same Qwen model scored 75% or 56% on larger overhauls depending on its agent harness, and model rankings changed with the environment. The benchmark does not include follow-up conversations or mid-task corrections, and its AI judges can err. Even with those limits, it suggests teams should evaluate a model and harness together on their own work rather than rely on nearly saturated leaderboards.