Moved an AI feature into production a few months ago and the cost profile has been a constant surprise since so the demos and the early prototypes ran cheap because the volume was tiny + the prompts were short but when it hit traffic the token usage scaled a lot. I think it was partly because custom
