#cascade

Try the small model first. If it fails or isn't confident, try the large one. Cascade routing gets 80% savings on 80% of requests.