OpenAI research identifies four training steps, including architecture choices and reinforcement learning, that make frontier models such as GPT-6 Astra harder to monitor through their chains of thought. It links the decline in transparency partly to decisions prioritizing capability and computing c