People will say that there is no reason to use Opus, unless you have too much money lying around. But I got to disagree. The gap between Sonnet and Opus is HUGE!
I personally use Opus only when I encounter issues that I and other "regular" model scan not fix. Most of the time these are infrastructure or build-step issues, not really application related. For example:
My local test pipeline for Coco started failing recently. The issue was an import that could not be resolved. I started debugging it manually and could not spot any issues. After 30 minutes I let Sonnet, Composer, Gemini and GPT at it. All of them took AGES (30 minutes +) and came to the conclusion that there must be a bug in Bun's package resolution logic. I knew this was not the case as the Github CI was still running fine.
Opus took 3 minutes to find a broken symlink that pointed to a newer version of the package that removed that particular export. But at the same time this single request cost about 5% of my regular monthly AI spending :D