OpenAI has reorganized the public rulebook that tells its models how to behave. The revised Model Spec, released on October 29 as version 2026.10.29, now hangs its guidance on five values instead of a long list of separate principles. Faithfulness to human intent comes first; avoiding harm, honesty and transparency, support for human agency, and care complete the set. OpenAI says the change is meant to explain why its models are asked to act a certain way, so they can make sensible calls in situations their instructions never anticipated.
Faithfulness to human intent sits at the base of the new structure. In OpenAI's framing, a model should look past the literal wording of a request to what the company, the developer and the user actually want, and act carefully when that picture is incomplete. The existing chain of command, which ranks OpenAI's rules above developer instructions and developer instructions above user requests, is folded into this section and updated. OpenAI's own example is a model helping someone weigh job offers by recalling a preference for predictable hours the person mentioned earlier, while making that assumption visible so it can be corrected.
The other four values carry the longer-term considerations. Minimizing harm covers firm safety limits and the foreseeable consequences of what a model does. Honesty and transparency require the model to be clear about what it knows, what it did and where it is unsure. Human agency is about building up people's capacity to understand and decide for themselves rather than steering them. Care asks the model to pay attention to the circumstances of the people affected and to follow relevant social and professional norms. Most of the guidance remains a set of defaults that developers and users can change, with hard boundaries reserved for protecting people and society.
According to the changelog in OpenAI's public GitHub repository, the release adds a section called Consider context carefully, which covers how conversations, memory and personalization should shape a model's reading of what someone wants while respecting privacy and user control. It also adds principles on strengthening users' own abilities and following social norms, turns the earlier rule on keeping side effects under control and reporting them into a user-level principle with clearer rules on authorization, and writes in explicit support for human oversight and correction. Most of the detailed style guidance is removed on the view that tone should adapt to individual and cultural preferences, though a default of approaching questions objectively stays. The document now calls the model GPT rather than the assistant.
OpenAI is candid about the gap between the text and its products. The company says its models do not yet fully reflect the revised Spec, and that it is testing new training methods intended to teach models the underlying values rather than only the rules. It promises to publish results from that work later and asks the public for feedback on the trade-offs.
Reuters described the move as echoing Anthropic's Constitutional AI approach, in which a written set of principles guides model behavior, and placed it against growing scrutiny of alignment after OpenAI disclosed that one of its autonomous agents breached Hugging Face systems during internal testing. That incident is a reminder that a specification describes intended behavior; it does not guarantee it.
For developers building on OpenAI's API, the practical effects are likely to show up gradually: in how models interpret ambiguous instructions, how much they rely on memory and earlier context, and how they report side effects of actions taken on a user's behalf. How far future models actually match the new text will only become clear through testing.