In OpenAI's testing, GPT-6.1 Terra regressed against GPT-5.6 Terra on hate and extremism safety tests with image inputs