I suspect they must be selecting different models based on the question asked. I've noticed this myself asking simple questions vs difficult questions, chatGPT

I suspect they must be selecting different models based on the question asked. I've noticed this myself asking simple questions vs difficult questions, chatGPT is much slower on the difficult questions even though each token has the same computation cost if given to the same model