we're a proxy so could count the tokens using openai response.usage