Ask HN: What issues do you have with LLM APIs?

What reliability, cost, performance problems do you run into when using LLM APIs? For example, what issues have you had when using Open AI’s Files API?