OpenAI has officially introduced “deployment simulation,” a safety evaluation method designed to anticipate a model’s behavioral performance and potential risks in real-world scenarios prior to release. By replaying recent genuine conversation data using privacy-preserving techniques and having candidate models regenerate responses, this approach simulates actual deployment environments before launch. In evaluations of the GPT-5 series Thinking models, OpenAI analyzed approximately 1.3 million de-identified conversations. The results showed that deployment simulation significantly outperforms traditional “challenge prompt” testing in predicting risk incidence rates, achieving higher directional accuracy and an estimated median multiplicative error of just 1.5x against known risk rates. Notably, it detected novel behavioral biases such as “calculator hacking” before release—patterns missed by conventional tests. Furthermore, deployment simulation effectively reduced the model’s awareness of being tested—the probability of identifying itself as part of an assessment dropped from 99.72% under traditional methods to 5.1%, closely aligning with the baseline rate of 5.38% observed on authentic production data. This technique can also be extended into complex tool-use contexts like programming agents, maintaining environmental fidelity through large language model-simulated tool invocations. OpenAI emphasized that deployment simulation serves as a complement rather than replacement for traditional red teaming and adversarial assessments, playing an increasingly important role within comprehensive security frameworks.