But usually it doesn’t really work like that, at least not for the cases I use it for. And chatgpt is even worse: it’s a conversational AI, so I burn 1000s of tokens fine tuning a response which will be slightly different than the previous response. With gpt3 I have the same issue: my products send a slightly improved prompt, based on the previous result to gpt3 and the new result improves. If you want automated quality, it’s rapidly becoming an impossible business model where humans or custom written software tools are simply (much) cheaper. And your clients know this because your competitors will offer them. It needs to become 0.001/1000 or less. It will do in a few years. It does, like you say, depend on your case. Many cases don’t need this much and there it is cheap, but those cases will turn out to be more demanding when the competition flares up with better results. You can easily burn a million tokens just playing around on an automated system connected to gpt3. Not many people want to pay $20 for just playing and testing things out.

The same happens with stable diffusion by the way; you need to generate many slightly different (improved) images from the same seed, maybe with inpainting to get to the final piece. That costs many tokens. High token prices as they are lead to (much) worse quality of the end result in sd and gpt.