I think fine-tuning is not getting enough attention, it's much cheaper/quicker to train and enables the model to understand the specific challenges more in-depth. It's more annoying to set up, but here's one of the first services to streamline it: finetunedb.com