OpenAI reorganizes its Model Spec around five values

OpenAI published a reorganized version of its Model Spec on 29 October 2026. The public document describes the desired behavior of its artificial intelligence models. The revision is organized around five values and clarifies guidance on context, oversight, and humans’ ability to make decisions and act.

These values encompass faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care for affected people. They bring together commitments already present in the specification. OpenAI says the revision also changes some default behaviors.

The chain of command has five levels, in descending order of priority. They are named Root, System, Developer, User, and Guideline. OpenAI’s release notes show that this hierarchy already existed in September 2025.

OpenAI sets the boundaries within which applications operate. Developers control their applications within that framework, while users direct the model’s work under those constraints. Many of the default rules can be changed by the developer or user, subject to higher-priority instructions.

OpenAI acknowledges that its models do not yet fully reflect the specification. The public document provides researchers with a reference point. Independent evaluations could compare models with previous versions on their autonomy, the visibility of their assumptions, and the ability to correct them.

To illustrate the use of context, OpenAI gives the example of comparing job offers. The model could take into account a previously expressed preference for predictable hours. The company says it should make that assumption visible and easy to correct, so that the person retains control over their choice. An incorrectly inferred preference could otherwise influence their decision.