Some AI company needs to create a model that can delegate simple tasks to 'stupider' models. I often encounter a task which is complicated enough to require a strong model like Opus, but which subdivides into a number of tasks - the vast majority of which are enough that 3.5 Sonnet could pick it up. All Opus would need to do is subdivide the task into easy and hard bits, then spin up a bunch of 3.5 Sonnets for the easy stuff.
This seems like such an obvious idea that I'm sure everyone is already working on it!
Claude code does utilize both the full Sonnet model and the lighter Haiku model in an automatic way. When you exit a Claude code session, it gives you the stats (tokens, cost, etc). I expect there’s a way to get this in-session as well.
This seems like such an obvious idea that I'm sure everyone is already working on it!