Co-Design Bench finds coding agents finish only a quarter of real web-app tasks
What happened
On November 8, The Decoder reported that Alibaba and several Asian universities built Co-Design Bench, which asks coding agents to modify existing web apps from real user requests, with two AI judges checking whether the features actually work. Of the ten models tested, the best, GPT-6 Astra, fully completed only 26.4% of tasks and met about 70% of requirements on average, leaving nearly three quarters of tasks needing human finishing. All models did worse when tasks mixed several kinds of edits, and the same model performed noticeably differently across execution environments. The team recommends reporting results for both the model and the environment.
Written by AI from the coverage · updated yesterday
Coverage
Follow the reports to see the story from different sides.
- The DecoderAI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
阿里巴巴与多所亚洲高校构建 Co-Design Bench，测试的十个模型中，表现最佳的 GPT-6 Astra 仅完整完成 26.4% 的任务。该模型平均满足约 70% 的需求，测试要求智能体根据真实用户请求修改现有网页应用，并由两名 AI 裁判检查功能是否实际可用。任务混合多种修改类型时所有模型表现均下降，同一模型在不同执行环境中的表现也有明显差异，研究团队建议同时报告模型与执行环境。
Heat over time
Heat now 5·Comparable peak 10（Nov 8 18:00）·Comparable change over 24 hours -49%
The trend only compares accounts observed without gaps, so its range may be smaller than the current heat. Hover or tap the chart for each hour; the left and right arrow keys step through it.