What do people think is the best Anthropic model for balancing cost/speed/performance for Claude Code?

Maybe this is weird but what I have found works for me is planning and drafting Claude code tasks completely outside of Claude code, this helps me bound the limits of each prompt to get reliable performance. I then /clear everytime before entering so I always get clean context. I have been using an earlier version of Claude Code and sonnet 4.6 for a while now because I am scared if I update or change model it might actually be worse. Either slower to work, more costly on tokens or less predictable behaviour. What set-ups do you use? Have you noticed drop or improvement in performance after updating Claude Code version? How do other Anthropic models compare for running Claude Code. Of course I expect the higher end models to have more capability in general, but they also likely do more reasoning etc and I prefer Claude Code to be execution focused personally.