OpenAI puts a 20% compute cost on its new AI safety monitoring

OpenAI halted reinforcement-learning training on its newest deployment-bound models for two weeks, hardened and red-teamed the systems, and limited the largest frontier run while executing smaller jobs to test safeguards. The new safety-monitoring system, covering all tool-enabled training and inference for models at GPT-5.6 Sol level or higher and all Astra inference, adds about 20% to compute cost, issues alerts within 30 minutes, and stops runs that remain uncleared.