Alibaba 发布 Co-Design Bench：最强模型也只能完整解决 26.4% 的任务
AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
Alibaba 搞了个新基准，专挑真实翻车任务来测，最强的 GPT-6 Astra 也只过四分之一，跟旧基准 99% 的成绩反差很大。
Alibaba 推出编程智能体基准 Co-Design Bench，测试集全部来自真实用户请求中曾让智能体失败的任务。表现最好的 GPT-6 Astra 完整解决率只有 26.4%，意味着约四分之三的任务仍需人工收尾。而同样的模型在旧基准上的成绩超过 99%，两套数字之间的落差暴露了现有评测与实际使用之间的差距。
AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test
Alibaba's new coding-agent benchmark, Co-Design Bench, tests AI agents on real user requests that already tripped them up in practice. Even GPT-6 Astra, the top performer, fully solves only 26.4 percent of tasks. On older benchmarks, the same models score above 99 percent. The article AI coding agents still need human cleanup on nearly three out of four tasks in a real-world test appeared first on The Decoder .
