OpenAI rewrites its Model Spec around five core values and now calls the model 'GPT'

OpenAI published a new version of its Model Spec on October 29, the public document in which the company sets out how its models are supposed to behave. Rather than presenting a long sequence of rules, the revision groups its guidance under five core values. It is the second update this year, after an August version that focused on interactions with teenagers.

Five values, one foundation

The five values are faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care. OpenAI treats the first as the basis for the rest: the model is expected to work out what OpenAI, developers and users actually mean rather than apply instructions literally. The Spec adds that the list is not meant to mark the limits of the human values the model should respect.

The structure built around those values keeps the existing chain of command, in which OpenAI sets the outer boundaries, developers operate within them and users direct their own work. Most rules are framed as defaults that developers or users can change, with firm restrictions kept only where OpenAI considers them necessary to protect people. The oversight section still says the model must stay corrigible, meaning open to legitimate human correction and control.

What changed

The update adds a subsection on considering context carefully, along with guidance on oversight and agency, and new defaults on supporting human agency and on controlling and communicating side effects. Most of the detailed style guidance from earlier versions has been retired, so tone is expected to follow user and cultural preferences. One stronger default survives: the model should approach questions from an objective standpoint, and only adopt a particular perspective when someone explicitly asks for it, not because it guesses what they want to hear.

The document also stops calling the model "the assistant" and refers to it as "GPT" throughout, which OpenAI links to the growing range of settings in which its models are deployed. An analysis by FourWeekMBA counted the number of headings falling from 81 to 70, while overall length stayed almost unchanged.

Limits and open questions

OpenAI acknowledges that the models it ships today do not yet fully match the new Spec and says it is experimenting with training methods meant to teach the values behind the rules, not just the rules themselves. The company is inviting public feedback on where defaults and hard limits should sit. How closely deployed systems follow the document remains something outside observers will have to test for themselves.