OpenAI Publishes Deployment Simulation: Pre-Deployment Safety Method Replays 1.3M Real Conversations to Catch Model Misbehavior Before Release

OpenAI published Deployment Simulation on June 16, a pre-deployment safety framework that replays approximately 1.3 million de-identified user conversations through a candidate model before release to surface misaligned behaviors before they reach production. The method analyzed sessions spanning GPT-5 Thinking through GPT-5.4 across August 2025 to March 2026 and extends risk assessment to agentic coding via simulated tool calls. OpenAI framed the approach as shifting bad-behavior discovery from post-launch auditing to pre-release stress-testing, complementing its existing post-deployment auditing programs.