GPT-5.6 Sol vs. Claude Fable: Why OpenAI’s new model crushes my benchmark