OpenAI has published a new version of its Model Spec, the document that defines the expected behavior of its models. It is based on five structuring values, gives greater importance to the context of exchanges, and places more emphasis on human intent.
On October 29, 2026, OpenAI unveiled a significant revision of this public framework, which governs how its models should respond to users. The document evolves as systems gain capabilities and autonomy. The stated goal is to go beyond simply executing instructions, so that models better interpret human expectations, including in novel situations. The change concerns safety as well as transparency and the control left to users.
Governing model behavior
The new architecture is based on five principles: fidelity to human intent, harm reduction, honesty and transparency, human agency, and care.
The first value is central. OpenAI wants its models to better understand what users, developers, and the company are actually seeking, rather than mechanically following instructions. This distinction comes as complex uses and systems capable of taking actions proliferate.
The four other principles complete this framework. They concern risk prevention, communicating limits and uncertainties, respecting users’ decision-making capacity, and considering people affected by the generated responses. This structure should also help models arbitrate between sometimes contradictory requests. However, respecting human intent remains subject to safety limits and human supervision.
Context and memory given greater consideration
Another notable evolution, a new section specifies how previous conversations, memories, and expressed preferences should contribute to interpreting requests. A model will thus be able to take into account information already provided by a user to adapt its response.
OpenAI nevertheless insists on the need to make these assumptions visible and easily modifiable, to prevent personalization from turning into implicit steering of choices. The revision also strengthens principles related to controlling the side effects of model actions. It introduces recommendations on respecting social norms and on maintaining users’ ability to understand, choose, and act for themselves.
Fewer rules on response style
OpenAI is also using this overhaul to remove a large part of the detailed guidelines dedicated to style. Writing behavior should now depend more on user preferences, context, languages, and cultures.
The principle of objectivity remains. A model must not infer that a user wants to hear a particular opinion simply because it thinks that opinion would suit them. A specific perspective can, however, be adopted when explicitly requested. Another change: the document now uses the term GPT rather than ‘assistant’, to account for the diversity of model uses.
This new version is primarily a behavioral framework, not a guarantee that all models already apply these principles. OpenAI acknowledges that current systems do not yet fully reflect the document’s directions. The company is also working on new training methods to improve their ability to integrate these values and apply them in novel situations.
This initiative comes as models capable of acting more autonomously are being developed. It also recalls work being done by other players, notably Anthropic with its Constitutional AI approach.
