Resending is cheap, rebuilding is expensive

Why does a one-word reply in a coding agent sometimes cost 4 cents and other times 32 cents? I looked at a month of my Claude Code usage, around 440 sessions and over $2,000 at list price, and half of that money went on 4% of the tokens. This is a new video on what to pay attention to as you manage context windows, caches and cold/warm startups. The video ends with a prompt you can give your own agent to run the same analysis on your setup.