On October 29, 2026, OpenAI published a revised public Model Spec, which describes intended behavior for GPT models in OpenAI products, including the API. The update organizes its guidance around five values and sets out changes to context, oversight, behavioral defaults and style guidance. OpenAI’s announcement and the dated Model Spec also explain how instruction authority affects customization.
The five values behind the Model Spec
OpenAI names faithfulness to human intent as the foundation of the framework. The other values are minimizing harm, honesty and transparency, facilitating human agency, and acting with care.
The Spec says explicit instructions are usually the clearest indication of what a person values. Context and prior interactions can help interpret those instructions. Its guidance on human agency calls for supporting people’s ability to understand and decide, rather than substituting the model’s prediction of what a person should choose.
What OpenAI changed in the document
OpenAI said the revision adds guidance on context, oversight and agency, adjusts some behavioral defaults, and retires much of the detailed style guidance. It also uses “GPT” instead of “the assistant,” a terminology change OpenAI links to the range of use cases and deployment settings for its models.
These are changes to the public framework’s guidance and terminology. They describe how OpenAI wants its models to behave, not a guarantee that every model already behaves that way.
How instruction authority limits overrides
The dated Spec names six levels: Root, System, Developer, User, Guideline, and no-authority-by-default. When applicable instructions conflict, higher-authority instructions take precedence. The levels help explain why saying that many rules are defaults does not mean every rule can be overridden by a user.
Override conditions vary. Guidelines can be overridden implicitly by context. User-level defaults require an explicit instruction at the user level or higher. Developers can override some defaults that users cannot, while higher-authority instructions remain in force when they apply. Text assigned no authority by default does not carry instruction authority merely because it appears in a conversation.
What OpenAI said about production models
On October 29, OpenAI said its production models did not yet fully reflect the Spec and that it was exploring training methods to help models learn its values and apply the principles in unfamiliar situations. The revised document sets out intended behavior; OpenAI’s statement distinguishes that goal from the behavior of its production models at the time of the announcement.
