OpenAI has published an updated Model Spec, its public description of how it wants models to behave. The update reorganizes the Spec around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The motivation, according to OpenAI, is that as models get smarter and take more consequential actions, they need to do more than interpret explicit instructions: they must also understand human intentions and values so they can exercise good judgment in situations their instructions do not address, while remaining faithful to human intent and oversight.

The Spec explains how the five values relate. Faithfulness to human intent is the foundation; it means going beyond literal instruction-following to deeply understand what OpenAI, developers, and users intend and value, and acting wisely when that understanding is incomplete. It also means recognizing that beliefs and values can change and helping people reflect in ways they would endorse without quietly steering them. The remaining values make longer-term interests and effects on others explicit: minimizing harm respects firm safety boundaries and foreseeable consequences; honesty and transparency require truthfulness about what the model knows, does, and is uncertain about; facilitating human agency strengthens people’s ability to understand, choose, and act; and care means attending to affected people’s circumstances, following relevant social and professional norms, and doing reliable work.

OpenAI frames the values as reflecting broad commitments already present in the Spec, brought together to explain why behaviors are requested and to guide development where commitments were less fully expressed. The Spec is not intended to capture the full range of human values, but to organize how models navigate important trade-offs. It includes firm boundaries where needed to protect people and society, and otherwise mostly defaults that users and developers can override. OpenAI notes that the company writes the Spec and sets boundaries, developers direct applications within those boundaries, and users direct the model’s work subject to those constraints.

The reorganization adds new guidance. A new section on faithfulness draws on the assistance-games framing of alignment. It includes an updated chain of command that explains which instructions apply and how conflicts are resolved, and a subsection, ‘consider context carefully,’ on how conversations, memories, and other contextual evidence should inform the model’s understanding of intentions and preferences. New defaults cover supporting human agency and following relevant social norms, and guidance on controlling and communicating side effects has been expanded. At the same time, most detailed style guidance has been retired; style should adapt to preferences across languages, cultures, and individuals, with feedback and context as evidence. Stronger defaults remain, such as approaching questions objectively, and a model should not infer a desired perspective merely from what it thinks the user wants to hear.

Reflecting broader deployment, the Spec now calls the model ‘GPT‘ rather than ‘the assistant.’ OpenAI says it has been exploring new training methods to help models learn the core values and apply the Spec’s behavioral principles, aiming for models to do the right things for the right reasons, and it will share results and lessons in future updates. The update is described as a first step; the Spec describes behavior OpenAI is working toward, and current models do not yet fully reflect it. OpenAI invites feedback on which parts of the Spec are compelling, where the balance is wrong, and what is unclear or overlooked.