I actually asked Claude about that this week and it was fairly interesting: the estimates are based on historical data, so I think that suggests in the current set up there will always be wild gaps. As best I could tell, the estimates you get are probably for the model one or two whole versions ago?