I don’t see how they could credibly estimate inference costs without knowing the model size.