I Tracked Every LLM API Call For a Week — 65% Were Unnecessary

I've been using GPT-5 and Claude via API for coding tasks — refactoring, code review, architecture questions, debugging. The bill was creeping past $150/month and I had no idea which calls were actually worth the money. Provider dashboards show you totals. Tokens used, dollars spent, done. But they don't tell you which specific calls were unnecessary. Was that $2.80 request for "where is the auth middleware" really worth sending to GPT-4o? So I built a tracker to find out. The experiment I wrote