few-shot learning requires providing a number of examples in the prompt making it longer

different capabilities emerge at a different scale - it's hard to tell which will suit your needs without running some tests first

if subsequent requests are based on previous model responses, the errors compound, so it is better to use the most powerful model, e.g. when building sophisticated autonomous agents