Your tests can make 66,000 paid LLM calls a month before your first user

A suite of 100 tests that call the model, run 30 times a working day, makes 66,000 paid calls a month. That's $290 on GPT-5.6 at list price and $528 on Claude Opus 5.5. Record each answer once, replay it on every other run, and the bill shrinks to one live run a night, under 5% of the calls.