Well it is mostly the GPT-4 that is expensive to run and because we generate 3 version of your long prompt, with it's inferences and our own prompt.