must be something wrong with the benchmark, the thing everyone optimizes for. That's actually a big red flag, and very cringe that you'd naively believe OpenAI.