Predicting model behavior before release by simulating deployment

OpenAI's deployment simulation tests a candidate model on recent real conversation contexts by replacing an earlier assistant response. This estimates how often unwanted behavior might occur in ordinary use, where questions differ from a deliberately difficult safety test. The study used de-identified conversations from users who permitted model-improvement use.

Across the studied deployments, predictions had a median multiplicative error of 1.5: an actual rate of 10 cases per 100,000 messages could correspond to a prediction of 15 or about 6.7 per 100,000. Some errors were much larger. The method complements monitoring long-running agents, but sampled traffic still leaves very rare failures to targeted testing.

“The main technique behind Deployment Simulation is simple: we take recent conversations from deployment, remove the original assistant response from the older model, and regenerate it with a candidate model to be released.”

OpenAI

“In total, we analyzed approximately 1.3 million de-identified conversations across GPT‑5 Thinking through GPT‑5.4 deployments, spanning August 2025 to March 2026.”

OpenAI

“Calculator hacking is a form of reward hacking which involves the model using a browser tool as a calculator while presenting the action as a search.”

OpenAI

“For example, GPT‑5.2 labeled real production traffic as evaluation-like 5.4% of the time and simulated traffic 5.1% of the time, while traditional evaluations (such as SWE-Bench or MonorepoBench, which are composed primarily of synthetic examples) were classified as evaluations almost 100% of the time.”

OpenAI

“In a test where a discriminator tried to tell simulated from real rollouts, realism improved from an 11.6% win rate to 49.5%, near the 50% chance level.”

OpenAI