Deployment Simulation: OpenAI Uses Real Distribution to Predict Model Risks Before Release
OpenAI released a Deployment Simulation paper in June 2026, using de-identified historical conversations to regenerate responses for models pending release, and comparing the simulated frequency with the real risk frequency after deployment.
Traditional safety evaluations tend to use manually selected difficult samples, making it hard to estimate the incidence of risks in real traffic. Deployment Simulation introduces production distribution into pre-release validation, shifting model risk assessment from whether it fails to how often it actually fails.
The study replays privacy-processed conversation prefixes on multiple generations of GPT-5 Thinking deployments, evaluating the direction and frequency calibration of 20 categories of undesirable behaviors, and extends to agent trajectories involving tool calls. The official disclosure mentions approximately 1.3 million de-identified conversations used for validation, while explicitly stating that risks rarer than about 1 in 200,000 cannot be measured.
Frontier model release gates are shifting from static benchmarks to continuous simulations close to production distribution; platforms with real traffic and robust privacy governance will gain an advantage in evaluation data.
Before releasing high-risk agents, enterprises should use real but de-identified historical tasks for shadow replay, compare failure rates between old and new systems, and retain dedicated stress tests for low-frequency, high-loss risks.
External audits are needed for privacy processing, sample representativeness, and grader bias, and to verify whether historical traffic can still predict new risks when product form or user base changes.