I'm pretty sure the model is cached with the system prompt already processed. So you should only pay extra tokens.