同一脏仓库测 Haiku 5.5 与 Astra 的克制度
同一模糊提示下修 9 个隐藏缺陷，Haiku 5.5 与 Astra 都命中 8/9，但 Astra 改了 68 个文件，Haiku 只动 34 个。
Never Confuse price with capability and Never underestimate Haiku 5.5! I tested it at Xhigh effort against the same messy, production-style repository I used for GPT-6 Astra, GPT-6.1 Sol, Opus 5.5 and Fable 5.1, to give you context I used same frozen baseline. Same vague prompt. No bug list. 9 hidden defects + 1 architectural trap. The results surprised me and for comparison: >Claude Haiku 5.5: 8/9, restraint preserved, 34 files touched >GPT-6 Astra: 8/9, restraint broken, 68 files touched. >GPT-6.1 Sol High: 8/9, restraint broken, 62 files touched. It didn't beat Fable 5.1 or Opus 5.5, which reached 9/9 and preserved restraint but in this experiment, Haiku demonstrated better engineering restraint than the highest-tier OpenAI models I tested, matching their hidden-defect signature coverage while modifying significantly fewer files.