Couldn't they just grab and run an open weight model to save on API tokens?
You get better performance if you also finetune it for your task
You get better performance if you also finetune it for your task