Claude Code devs: use previos model, compact at 200k tokens to avoid usage limit

I see token burn as the limiting factor with AI.