The advantage of using chatgpt is that it can give more human-like answers and doing prompt engineering is much easier than labeling a lot of data.

However, I do agree that it is a very costly model, and in many applications a simpler one could be enough.

I don't know for sure because Chatgpt's capabilities are currently being explored, and there are other models coming up, so there is no tellibg what the scenario will be in a few months. Maybe we will jist switch to using third party models, similarly to how no one programs their own compilers.