Michael JohnsonOn October 29, OpenAI updated the Model Spec, a public description of how the company wants to train its models to behave. The new revision organizes guidance around five values and expands rules for situations in which models use context and act with tools. For developers, it is primarily a document about intended model behavior, not an announcement of a new API or ready-made feature.
The updated Model Spec lists faithfulness to human intent, minimizing harm, honesty and transparency, supporting user autonomy, and care. OpenAI says these principles should help models make more appropriate decisions when a specific instruction does not cover every circumstance.
Context and oversight become part of the rules
The practical change concerns how a model should take interaction history and other information about the user into account. As an example, OpenAI describes comparing job offers: a model might consider a previously stated preference for a predictable schedule, make that assumption explicit, and give the person a chance to correct it. The document also includes more explicit guidance on human oversight, user autonomy, and the consequences of model actions.
OpenAI describes the specification as a set of boundaries and defaults that can mostly be overridden: developers can set the behavior of their applications within these limits, while users can guide the model’s work. For product teams, it is a useful reference when designing system instructions and checks: determine what context may be considered, how the model communicates its assumptions, and when a decision requires human involvement.
Fewer style prescriptions
In the same revision, OpenAI said it is removing most of its detailed style guidance. The company explains that style should adapt to people’s preferences, languages, and context; more enduring requirements include basics such as answering questions objectively. This shifts the specification’s emphasis from a single set of stylistic rules to values and ways of resolving ambiguous situations.
OpenAI also says it is exploring training methods to help models apply these principles more effectively, and promises to share results in future updates. At the same time, the company explicitly describes the specification as a statement of intended behavior: current models do not yet fully match it. Teams adopting it in products should therefore use the document as a basis for requirements and test specific scenarios with their own evaluations of the application’s tasks, context, and tools.
The main practical takeaway for teams is to review instructions and tests not only for whether the model follows requests accurately, but also for how it uses context, discloses assumptions, and keeps humans in control. The update was published on October 29; OpenAI did not announce a separate model release or new interface in this announcement.
