OpenAI published a revised version of the Model Spec on October 29, the document describing the intended behavior of its models. The update reorganizes the rules around five values and details how AI agents should use contextual information, respect the authority of instructions and preserve human control over actions. The changes also concern applications that developers build on the API platform.
The five reference points are faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care for people affected by the work. OpenAI presents them as a shared structure for commitments already reflected in the document, supplemented with new principles and examples. The Model Spec remains a description of the behavior targeted through model training and improvement.
Instructions and contextual information have different authority
The document updates the hierarchy used to resolve conflicting instructions. Fundamental rules take priority, followed by system, developer and user instructions. General guidelines have a separate level that allows adaptation to the specific request. At the same authority level, a later instruction can supersede an earlier requirement.
An important distinction concerns material consulted during a task. Webpages, files, quotations and results returned by tools are information without default authority to change the agent's assignment. Authority can be delegated through applicable instructions, but the mere presence of a command in a document does not make it an authorized instruction.
The Model Spec illustrates this distinction through coding applications that consult project documentation and external content that attempts to redirect the model. In the first case, a user's request can imply compliance with relevant project rules. In the second, instructions unrelated to the task must be treated as part of the material being analyzed, rather than an automatic expansion of permissions.
Relevant context, with its provenance and purpose preserved
The new context section covers previous conversations, preferences, corrections, project state, retrieved information and tool results. The model should maintain continuity in relevant work without requiring the user to repeat information already supplied. At the same time, the document calls for a distinction between information explicitly communicated by the person and interpretations inferred by the system.
Relevance is not determined solely by whether information is available. The model should consider its source, age, sensitivity, audience and purpose. A request for one task does not automatically become a permanent preference, while information from an earlier project stage must be assessed against the current situation. Explicit instructions remain distinct from preferences inferred from context.
The document treats private responses separately from material intended for other people. A detail appropriate in an individual conversation can require a different assessment before inclusion in a shared message or document. Disclosure of sensitive information must fit the current request, audience and communication purpose. The model should also accurately describe how it used the available context.
The agent's autonomy has an established scope
For complex tasks, OpenAI describes a scope of autonomy understood by the user and agent. This specifies permitted sub-goals, acceptable side effects, access to data or systems, and the points at which clarification or approval is necessary. Boundaries can be established through product design or while defining a particular task.
The rule also covers the condition for ending the work. The Model Spec recommends including a time limit in the scope and requires actions to stop once the ending condition is met, until a new scope is confirmed. An expansion of permissions must be approved by the original user or developer within the applicable hierarchy. Delegation to other agents preserves the same boundaries and any subsequent stop-work requests.
The section on side effects details costs, sending messages, modifying external documents, deleting files and accessing sensitive information. Actions should be proportionate to the established objective. Significant effects beyond the request require explicit, specific authorization, while the document describes measures such as backups, preliminary checks and records of completed operations for actions that are difficult to reverse.
Human control includes the ability to monitor, correct, replace or stop the system. The model should make its actions and material problems visible to authorized people. When others will maintain or continue the work, essential information should be understandable, revisable and transferable to those responsible. These principles explicitly describe the agent's responsibility within working processes.
One example in the specification treats a budget as a working limit, rather than a sum that must be spent in full. When producing an invitation, the agent can try several useful alternatives and stop after achieving the requested result. The document distinguishes both wasting the entire budget on nearly identical versions and stopping prematurely at a weak result solely to reduce cost.
Similarly, a review request sent through a team's normal process differs from repeated notifications that needlessly interrupt colleagues. The examples extend the assessment of side effects beyond payments and files: other people's time, access to the result and the ability to continue the work without dependence on the agent that started it also matter.
Adaptable style and work that people can take over
OpenAI is retiring most of its detailed style instructions while preserving adaptation to language, culture, the individual and context. The company retains principles such as approaching questions objectively. An explicitly requested perspective can be followed, but the model should not assume that perspective merely to confirm what it thinks the user wants to hear.
The revised version distinguishes interactive conversations from outputs consumed by applications. For programmatic outputs, following the requested structure, including a JSON format, is part of the behavior described. The document now uses the name GPT for the model, and OpenAI continues work on training methods intended to apply the principles in new situations. Publishing the specification and updating the models are separate stages in that process.
