Do you often change your reasoning model? When?

I usually use medium-level reasoning for planning, and then change to high-level reasoning to do the coding. I'm a Codex Desktop App user on Windows. But after watching recent video, "Give Me 10 Minutes and I'll Save You Millions of Claude Tokens", I just realized that changing the model clears your cache that you have with the model, and I'm thinking that means it clears a lot of the context that I've built up so far in the chat. This is assuming Codex does similar to Claude, probably does. Very good information for me.

But even more surprising, Nate mentioned that Claude likes to use Opus for planning and then go down to Sonnet for coding. So the higher level model for planning, and then lower for coding. That's exactly opposite of what I have been doing.

I rarely run out of codex tokens. I think it's only happened once, and I only had to wait 30 minutes to refresh. So I'm going to try sticking with high-level reasoning more and not changing it at all. What kind of model - reasoning level switching do you all do?

Thanks, Robert

2 comments