Also to elaborate on what I meant earlier, maybe you could "cache" GPT results and reduce costs by letting users publish the better GPT results on your service as templates.