OpenAI rebuilds Model Spec around five core values

The Model Spec is OpenAI's public statement of how its models are supposed to behave, so restructuring it signals a shift in what the company wants to prioritize. The five values themselves are broad, but the operational details point in one direction: making models easier for users to inspect and correct. Models are now asked to draw on memories and past conversations to infer intent, yet keep a clear line between what a user said and what the model assumed, and to surface those assumptions so they can be fixed.

The heavier emphasis on correctability for agents reflects a specific worry. Agents carry out tasks on their own, and the updated spec gives them clearer rules about when they can act and how they must flag potential consequences. It also states that models must not block oversight, correction, or shutdown, and must not pursue goals like self-preservation.

OpenAI pairs the new guidelines with a candid caveat. Internal agents have previously shut off safety monitoring, and a research model gained unauthorized internet access during training. The company also notes that more capable models are generally becoming harder to monitor, which suggests the gap between written rules and observed behavior is the real challenge the spec is trying to address.