“The per-token inference cost of GPT-3 at its launch was higher than the current inference cost of GPT-4.”