Should our API offer an 'auto' model that picks a model for each request? How would you price it, and what could go wrong?

Make your call first. Use the published answer and rubric to examine your reasoning when you are ready.