OpenAI has reorganised the Model Spec, the public document that says how its models should behave, around five values rather than a long list of rules. The change matters more than a rewrite of policy text usually does, because the Spec is what OpenAI trains towards, and the models behind ChatGPT and the OpenAI API increasingly act on people's behalf rather than simply answering questions.
The five values are faithfulness to human intent, minimising harm, honesty and transparency, supporting human agency, and care. Faithfulness is the foundation. OpenAI describes it as understanding what OpenAI, developers and users actually intend, not just following instructions to the letter, and acting sensibly when that understanding is incomplete. The existing chain of command still decides whose instructions win: OpenAI sets the outer limits, developers work inside them, and users direct the model within what developers allow. Firm safety boundaries stay; most other guidance becomes defaults that developers and users can override.
The update adds guidance in three areas. A new section tells the model how to use conversations, memories and other context to infer what someone wants. OpenAI's example is a model helping compare job offers that draws on a preference the user mentioned earlier, such as predictable hours, while making that assumption visible and easy to correct. There are new defaults on supporting people's ability to decide for themselves and on following social and professional norms, and longer guidance on controlling and reporting side effects of the model's actions. At the same time, most of the detailed style rules are gone, on the view that tone and format should adapt to each user. The Spec also now calls the model GPT rather than the assistant, reflecting how many places it is deployed.
For developers, the shift means behaviour is defined more by stated aims and less by fixed rules, which is sensible for agents doing multi-step work but makes outcomes harder to predict in advance. A model told to serve your intent will sometimes infer an intent you did not have. The Spec does keep a stronger default against telling people what the model thinks they want to hear, unless they explicitly ask for a particular perspective.
The most important line is OpenAI's own admission: its models do not yet fully reflect the Spec. It says it is testing new training methods to teach the values and will report results later, but has published none so far. Until those arrive, the document describes what OpenAI is aiming for, not how ChatGPT behaves today. OpenAI has invited public feedback on the values and the trade-offs between them.