It’s more like % better for X the cost (or time) to train/tune tradeoff . For example gpt4 is significantly slower and 15x more expensive than 3.5-turbo

AFAIK, we can't currently finetune 3.5-turbo and from my experience, prompts/consistent results in 4 vs 3.5 is like comparing a tool to a toy. (Although I know my focus on building internal tools vs end-user tools taints my view of "cost"...) Honstely, I just want gpt-4 to be faster and cheaper