From GPT-4 to Mistral 7B, there is a 300x range in the cost of LLM inference