I would be curious to see if the results improve by using DSPy to improve your prompts (and also reevaluate which prompts work better on the newest model).