Predicting model behavior before release by simulating deployment

OpenAI has introduced a method called Deployment Simulation that uses real conversation data to anticipate how a model will behave once it reaches users, before that release happens. The approach is positioned as an improvement to safety evaluation accuracy, moving beyond static benchmarks toward something more ecologically valid. How well simulated deployment captures the full breadth of real-world use remains an open and important question.