OpenAI resumed reinforcement learning training and internal inference for its frontier AI models on October 1, 2026, following a training pause in September 2026