Alibaba and university researchers introduced Co-Design Bench, a coding-agent benchmark built from real user change requests, and found that GPT-6 Astra, the best of ten models tested, fully completes only 26.4 percent of tasks, despite meeting about 70 percent of requirements on average. The researchers cite saturation of existing benchmarks, where models exceed 99 percent, and recommend reporting models and their harnesses together, since the working environment strongly influenced results.
No score is assigned. Sources and their independence are shown in the citation chain below.