Ok so add a factor of 20 to match the 8000 questions that OpenAI tested on, it's still crazy efficient compared to the NS result