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