Five values instead of a rulebook
OpenAI published an updated Model Spec on 29 October, reorganising the document around five stated values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. The company described the change as a step in work on value alignment, aiming for models that can apply human values in situations their instructions do not explicitly address.
The Model Spec matters because it is the only public text in which OpenAI commits to how its models should behave, and because the company trains against it. It is not a description of how the models do behave. OpenAI said so directly: the Spec describes the behaviour it is working toward, and its models do not yet fully reflect it.
What was added
The new section on faithfulness to human intent starts from the position that what people value comes first and that instructions are usually the clearest signal of those values. OpenAI said the approach borrows from the assistance-games framing of alignment, and that the section carries an updated chain of command setting out which instructions apply and how conflicts between them are resolved.
A new subsection, consider context carefully, covers how conversations, memories and other contextual evidence should shape the model’s reading of what a person wants. The company’s worked example is mundane and revealing: asked to compare two job offers, a model might draw on a preference the user mentioned earlier — predictable hours — while making that assumption visible and easy to correct.
OpenAI also added defaults on supporting human agency and on following relevant social norms, and expanded existing guidance on controlling and communicating the side effects of a model’s work.
What was removed
The larger change is a subtraction. OpenAI said it is retiring most of the Spec’s detailed style guidance, on the reasoning that style should adapt to what people prefer — in aggregate, within languages and cultures, and per person — with user feedback and context supplying the evidence.
One limit survives. The Spec keeps stronger defaults such as approaching questions objectively, and OpenAI drew a line around them: users and developers can explicitly ask a model to take a particular perspective, but the model should not infer that preference from what it thinks they want to hear. That is a sycophancy rule stated as a default rather than as a style note, which is the point of moving it out of the style section.
The model is now GPT
The document also stops calling its subject “the assistant”. OpenAI said it now refers to the model as “GPT” throughout, citing the widening range of use cases and deployment settings its models run in. A naming change in a behaviour document is not a product decision, but it does remove the one word in the Spec that framed the model as a helper standing beside a user rather than a system acting in an environment.
OpenAI said it has been exploring new training methods intended to teach models the five values rather than the individual behavioural rules, with the aim of getting models to do the right things for the right reasons, and that it will publish results and lessons from that work in future updates. The company invited public feedback on the values, the trade-offs between them and the specific language.
What to watch is whether the removal of style guidance shows up in deployed behaviour. The Spec now leaves tone largely to feedback and context, which means the next observable change will come from training data and preference signals rather than from a line anyone can read in the document.
