A benefit of smaller models: parallelism.

E.g. GPT4o-mini is so cheap and fast, you can run a number of different iterations on a prompt in parallel and then pick the one you like the best.

Or let the user pick if you have a 4-up evolution style UI.