同样的提问,不同结果

一次 Claude 检索对比

同一个 Project,几乎一样的提示词,让 Claude 的 Haiku 和 Sonnet 去搜索Project中同同样所有旧对话——结果差挺远。
Project 里有两样东西:多个上传的文件和多组聊天记录。每组对话有对话名称。Sonnet 分得清,直接去搜;Haiku 却一直翻文件,翻不到,就把文件里的标题当成”对话名称”塞给我——没找对地方,还不肯承认没有。我需要改提示词为“历史聊天记录”,Haiku它才能找到结果。
更麻烦的是,它最后那份”完整”清单也不可靠:思考摘要写”四个”,正文却写”五个”,自己都对不上;还有几条只有说明、没有正文,像硬凑的。Sonnet 则干净地给出两个,都带完整内容。
Haiku智商大概10岁，Sonnet已经成年。

Same question, different models, quite a gap
(a little Claude test)
Same Project, almost the same prompt. I asked two Claudes, Haiku and Sonnet, to dig up all the old conversations on one topic inside that Project. The results came out pretty far apart.
A Project holds two kinds of things: the files you upload, and the chat conversations — and each conversation has its own name. Sonnet could tell them apart, and went straight for the conversations. Haiku could not. It kept flipping through the files, found nothing, and then handed me a title pulled from a file as if it were the “conversation name.” Wrong place to look — and it would not admit it had nothing. Only after I changed my prompt to “chat history” did it finally find the right thing.
The worse part: even its final “complete” list did not hold up. The thinking summary said “four,” the answer said “five” — it could not even agree with itself. A few entries had only a description and no real content, as if they were dropped in just to fill the list. Sonnet, on the other hand, gave a clean two, both with the full text.
Haiku feels about 10 years old. Sonnet has already grown up.