OpenAI's Deployment Simulation predicts undesired model behavior before release by replaying past production conversations through a candidate model

OpenAI's Deployment Simulation predicts undesired model behavior before release by replaying past production conversations through a candidate model

OpenAI's Deployment Simulation replays 1.3M conversations to surface GPT-5 behavioral drift before release. How it caught calculator hacking. June 2026 guide.