NOTE: this is only valid for the old models (GPT-3 and Codex). IIRC, there is no simple way to know the token usage for the new models (gpt3.5-turbo and beyond).