HN 提问：您认为 Opus 5 会改进吗？
4 分•作者: logicallee•7 天前
Claude 的 Opus 5 是一个近期推出的模型，作为 Opus 4.8 的直接替代品，它带来了许多问题，包括忽略我们记录详尽的部署流程（在我们的 claude.md 文件和架构文件中都有清晰描述），并在激活后立即导致生产环境崩溃。
目前，我添加了一个监督层，由 Opus 4.8 实例监督 Opus 5 并对其进行纠正。这起到了一定的作用。我正在考虑降级回 Opus 4.8，但想再给 Opus 5 一次机会。
您认为这个模型会改进，还是会停留在当前的能力水平？
您可以看到其他人的抱怨：
https://www.reddit.com/r/ClaudeAI/comments/1v92csh/opus_5_extremely_rlfried_and_mistakeprone_for/
“在我看来，Opus 5 真的不适合任何任务。我在一个大型、复杂的项目中尝试了它各种可能的角色，它在所有任务上表现都很差。”
Claudebot 对该帖子的总结：
> 该帖子中的压倒性共识是，Opus 5 是一个有缺陷、过度自信且混乱的模型，对于复杂编码来说，相比 Opus 4.8 是一个重大的倒退。
查看原文
Claude's Opus 5 is a recent model and has been very problematic as a drop-in replacement for Opus 4.8, including ignoring our well-documented deploy process (clearly described in our short claude.md file and short architecture file) and breaking production immediately upon being activated.<p>Currently, I added a supervisor layer where an Opus 4.8 instance supervises Opus 5 and corrects it. It helped a little. I am thinking of downgrading to Opus 4.8 but want to give Opus 5 a chance.<p>Do you think this model will improve or will it stay stuck at its current level of capabilities?<p>You can see complaints from other people:<p>https://www.reddit.com/r/ClaudeAI/comments/1v92csh/opus_5_extremely_rlfried_and_mistakeprone_for/<p>"Opus 5 is truly not good for any task, imo. i have thoroughly tried it in every possible role in a large, complicated project. it is bad for all tasks."<p>Claudebot's summary of the thread:<p>>The overwhelming consensus in this thread is that Opus 5 is a buggy, overconfident mess and a significant regression from Opus 4.8 for complex coding.