AI reasoning models used 30 times more power on average to respond to 1,000 written prompts than alternatives without it, a new study finds.