Best practice for efficient credit use

I have read online that the best practice is always to use a high thinking model like Opus 4.7 to do the planning, and then we can use a simple model, like Sonnet 4.6 to do the implementation.

Or can we even use composer for this / auto mode to save us some credits cause probably stick to Opus 4.7 to get the best possible results will just burn credits, right?

Im asking because sometimes I have more difficult things to do like working with mcp servers or updating a n8n workflow using the rest api. I found the composer to be a bit dumb when I comes to doing these tasks since it always generates a lot of clutter files (e.g. scripts that deploy an edge function to supabase while opus directly runs it).

You guys are aware of any other downside to implement in auto (or premium with cheaper model)?