OpenAI tests unreleased models on replayed user conversations
OpenAI has added "Deployment Simulation" to its pre-release safety review, regenerating responses to past user conversations with candidate models to forecast how often undesired behavior will show up after launch.
OpenAI described a method it calls Deployment Simulation. The company takes recent conversations from deployment, removes the older model's assistant response, and has a candidate model generate a new one. OpenAI says the replay is done in a privacy-preserving way. It then searches the new completions for failure modes and uses their frequency to estimate how often those behaviors will occur in real use.
The company says the method works alongside targeted evaluations and red-teaming rather than replacing them. Traditional evaluations are still needed to stress-test rare, high-severity scenarios. OpenAI says the new approach can't be expected to measure behaviors that occur less than once in 200,000 messages.
According to OpenAI, the method deals with three weaknesses of conventional evals:
- Coverage: it gets broader by simulating more traffic, so it scales with compute rather than with the manual work of building new tests.
- Selection bias: the prompts reflect recent real usage.
- Test recognition: OpenAI says models don't seem able to tell simulated conversations from real traffic.
To validate the method, OpenAI pre-registered predictions of how often 20 types of undesirable behavior would occur once GPT‑5.4 Thinking was deployed. It also ran retrospective studies on other GPT‑5-series Thinking models. The categories include misalignment, such as a model lying about tools, and disallowed content. Because the same measurements can be run on live traffic after release, the forecasts can be checked.
OpenAI says the method improved its estimates of undesired-behavior rates and surfaced new forms of misalignment before release. It also says the method extends to agentic, tool-using rollouts and to risk assessment before internal deployments. The excerpt does not include accuracy figures for the pre-registered predictions.