The trick is managing the context (number of tokens). It’s tempting to have a long conversation about tweaking some code, but a shorter context (last X messages), or ‘forking’ the conversation regularly can both improve performance and drastically lowers cost.