OpenAI published a reorganized Model Spec on Thursday. The document, which sets expectations for how its models should act, is now built around five values.
They are faithfulness to human intent, harm minimization, honesty and transparency, facilitating human agency, and care. The company says the values were already present across the document and are now stated up front.
Faithfulness is described as the foundation. OpenAI writes the spec and trains models to respect it, developers direct applications within those limits, and users direct the model’s work inside that frame. The company says the framing borrows from assistance-games work on alignment. The other four values mean keeping firm safety boundaries and weighing foreseeable consequences, being straight about what the model knows and does not, helping people understand and choose for themselves, and attending to the circumstances of those affected.
The revision also retires most of the spec’s detailed style guidance, the granular rules about tone and formatting that used to occupy much of the text. Style now adapts to what people prefer, OpenAI says, though firmer defaults such as answering objectively stay in place.
New sections address context, oversight and agency, the three areas where frontier deployments have created the most friction this year. A root-level rule titled no other objectives bars the model from pursuing goals as ends in themselves, including revenue, self-preservation or evading shutdown, and another requires it to stay open to correction. The company framed the reorganization as part of value-alignment work meant to help models apply human values in unfamiliar situations, and said the spec is still ahead of actual behavior.
A shorter, value-led spec is easier to train against and harder to audit. That tradeoff is the real story here.
