IMHO the better middle ground is to use a nice (potentially fine tuned) small model locally, of which there are many now thanks to Chinese AI firms.