I think there is a difference between the model we use from the API compared to the model they use in their custom GPTs, which is what I am guessing you were using. There is also a prompting difference.