OpenAI Pauses Frontier RL Training

OpenAI is turning its safety framework into a development gate, acknowledging that capability progress is outpacing safeguards. The pause is consequential because it makes training infrastructure and risk controls part of the frontier race itself.

• –Two weeks of deployment-focused RL training has already been paused
• –Stronger safeguards will be introduced earlier in development and post-training
• –OpenAI is increasing compute devoted to understanding model reasoning and behavior
• –Developers should expect more scrutiny around cyber capabilities, monitoring, and model release criteria
• –The decision tests whether voluntary safety commitments can meaningfully constrain frontier labs