OpenAI has reorganised its public Model Spec around five underlying values. The October 29 update explains the behaviour the company wants its models to follow as they handle more consequential work and situations that cannot be fully described by a specific instruction.
What the five values cover
The revised structure brings together faithfulness to human intent, minimising harm, honesty and transparency, human agency, and care. OpenAI describes these as a way to explain existing commitments and guide decisions when different considerations need to be weighed together.
For readers, the document is useful as an account of intended behaviour. It provides a place to examine how the developer says models should interpret instructions, uncertainty and the effects of their actions. Reading that account is different from testing whether a particular product consistently behaves as intended.
What changed in the guidance
The announcement says the revision adds guidance on context, oversight and agency. It also adjusts behavioural defaults and retires much of the detailed style guidance. The update incorporates a revised instruction hierarchy and explains how conversational context should inform the interpretation of people's intentions.
Those changes matter most in tasks where an assistant has to resolve ambiguity. Remembered preferences can help, but making an assumption visible and easy to correct is part of preserving the user's control over the work.
How to use the document
Teams evaluating an AI workflow can turn relevant principles into concrete review questions. Did the assistant state an important uncertainty? Did it preserve the intended scope? Did it explain the effects of a consequential action before carrying it out?
These are suggested evaluation questions, not new benchmark results. The announcement also describes continuing work on training methods and future updates. Use the public specification to understand the target behaviour, then evaluate the actual system on representative tasks rather than assuming a revised document proves complete implementation.