“OpenAI's advice for the GPT-5.6 'thinking dial' is to start at the setting used on the previous model and then test one level lower.”