Notes from actually running things — mostly about keeping an AI agent alive on free API tiers without a bill that grows, and what holds up once it runs every day.
Free API tiers rotate weekly. A live route scout, a fallback chain and seven lessons from running a real agent without paying for the models — including the bug where my own cap silently dropped a working route.
I re-probed every free LLM API I hold a key for. Groq is alive, Gemini and Mistral are throttled, NVIDIA's IDs I tried are gone - with the method to re-run it.
Cost per task is not the token price. A method to measure what an agent really spends, and why routing matters more than the model.