OpenAI cuts inference costs in half with new optimization technique