Fine-Tuned Smaller Model or Few-Shot In-Context Learning Larger Model?

So far I've been using mostly few-shot in-context learning + recursive prompts on GPT3 da vinci. Decent results, but less "token-efficient" of course, because I have to prepend the semantic context or instruct format every time.
Fine tuning a smaller model (GPT-2 or GPT-J) for specific roles may be a better solution long term, or it may just be a waste of money.
For the AI veterans here, which method do you use?