First, there is no reason to pay double if you don’t need the context length of -16k. If it is really possible for you to exceed -4k with your application, you can make a more intelligent model-select mechanism in your software.
You should find a way to minimize the system prompt. The functions should have a clear description and names so they can stand alone on their own merit.
gpt-3.5 models have been hit with quality degradation for following system instructions going back a month. You cannot fix this. (They broke GPT-4, and thus had to go after gpt-3.5 that still worked and embarrassed the 30x more expensive model?)