I wonder whether it could be related to some kind of over-fitting, i.e. a prompting style that tends to work better with the older models, but performs worse with the newer ones.