Today I'm thinking: the GPT-4 API is about 30 times more expensive and significantly slower than GPT-3.5, yet it is noticeably better on complex questions (at least I see a fundamental difference in programming tasks, chemistry, and working with languages other than English and Spanish).
Hypothesis: based on the prompt and previous conversation (for example, when the user points out an error), we can choose which model to use and dynamically switch between them.
Question: how to do this effectively? Is there already something open-source?