Ask HN: Do you think Opus 5 will improve?
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.
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.
Do you think this model will improve or will it stay stuck at its current level of capabilities?
You can see complaints from other people:
https://www.reddit.com/r/ClaudeAI/comments/1v92csh/opus_5_extremely_rlfried_and_mistakeprone_for/
"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."
Claudebot's summary of the thread:
>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.
No comments yet.