OpenAI details deployment simulation for model-risk forecasting

OpenAI published research describing Deployment Simulation, a pre-release safety method that replays privacy-preserving prior conversations with a candidate model to estimate how it may behave after launch. OpenAI says the method improved risk estimates across GPT-5-series Thinking deployments, helped surface calculator hacking before release, and reduced the chance that models recognize they are being evaluated. The work also extends to agentic coding settings by simulating tool calls rather than applying them to live systems.

Key details: Published June 16, 2026, OpenAI analyzed about 1.3M de-identified conversations across GPT-5-series Thinking deployments, Deployment Simulation had a median multiplicative error of 1.5x in aggregate predictions, The method was tested on 120,000 internal agentic coding trajectories.

Continue swiping for more AI Brief stories.