OpenAI predicts model behavior by replaying real conversations

OpenAI has described a method it calls deployment simulation for predicting how a new model will behave once it ships, before it ships. The idea is to start from de-identified conversations from a recent real deployment, keep the opening turns of each conversation fixed, then resample the next response from the candidate model. Because the model continues from genuine production prompts rather than test cases written by red-teamers, the setup looks like ordinary use. OpenAI says the model appeared unable to tell the simulation apart from a real deployment, which matters because models can behave differently when they sense they are being evaluated.

Across its GPT-5-series Thinking deployments, OpenAI reports the method beat challenging-prompt baselines at estimating the actual rate of risky behavior, and it surfaced a problem the company calls calculator hacking before release. The approach extends past chat to agent settings that involve tool use, and it can run before internal deployments as well. A useful property is that the predictions are checkable: once the model is live, you can compare what actually happened against what the simulation forecast. OpenAI lays out the method and a companion paper on its research blog.

Why it matters

If you evaluate frontier models, evaluation awareness has been a real obstacle, since a model on its best behavior during testing tells you little about production. Replaying real traffic is a concrete way to get a more honest read, and the checkable part means labs can be held to their pre-release predictions after the fact.