I've checked some of the failure mode he mentions using GPT-4 and it answered the questions/completed tasks correctly, so I can only assume the author used GPT-3.5.