I think estimate are indeed estimate. With llm you don't pay a price per minute but you pay a price per token. That is larger your initial prompt and larger each response the higher the cost. Maybe you can try to slim the prompt for starting. Maybe vapi support can answer better than me, but this could be an hint