OpenAI's GPT-5 ditches refuse-or-comply safety training

OpenAI's GPT-5 uses "safe completions" — grading the safety of outputs instead of refusing based on inputs — and beats o3 on both safety and helpfulness for dual-use prompts.