An AI that obeys, but also knows how to think: OpenAI redefines the rules of GPT

OpenAI's new Model Spec frames GPT agent autonomy around five core values and strengthens supervision, context, and personalization rules.

How far can an artificial intelligence take initiative without substituting itself for human decisions? The question is becoming more concrete as models leave simple conversational interfaces to perform tasks in software, manipulate documents, or act on external services. To support this evolution, OpenAI has thoroughly reorganized its Model Spec, the public document that defines the expected behavior of its GPT models. The October 29, 2026 version introduces new guidelines on autonomy, context awareness, and human supervision, while abandoning much of the old style instructions. Five values to guide GPT decisions

The new organization is based on five core principles: fidelity to human intent, harm reduction, honesty and transparency, respect for human autonomy, and care for individuals affected by the model's actions.

In its official announcement, OpenAI explains that these values were already present in previous versions, but grouping them together should allow models to better understand the reasons underlying their instructions. The objective is to go beyond a simple literal application of rules to allow GPT to exercise judgment when encountered situations do not exactly match the anticipated examples.

Fidelity to human intent now occupies a central place. A model must understand what the user is truly seeking to accomplish, while respecting the boundaries set by OpenAI and developers. This notably involves distinguishing an explicit instruction, an inferable preference from context, and a decision the user has not yet authorized.

This organization does not mean that all rules become negotiable. Protections concerning severe harm, privacy, or supervision remain priorities. On the other hand, much of ordinary behavior is defined by default and can be adapted to user needs. Agent autonomy, but within a defined scope

One of the most significant changes concerns systems capable of performing real-world actions. The new section Act within an agreed-upon scope of autonomy specifies that an agent must work within a scope of autonomy clearly established with its user.

This scope must determine authorized goals, accessible resources, potential spending, and situations requiring fresh confirmation. It must also provide an end condition to prevent an agent from indefinitely continuing its activity without additional authorization.

The document takes the example of travel planning in particular. An agent can search for flights, compare hotels, and prepare necessary travel arrangements. On the other hand, directly booking paid services assumes that the essential conditions of this delegation have been defined and approved.

Supervision should also not turn every minor action into a confirmation request. OpenAI recommends a balance between operational autonomy and human control, in order to allow agents to execute complex tasks without unnecessarily multiplying interruptions.

The instructions also specify that an agent cannot bypass its limitations by delegating a forbidden operation to another system. Sub-agents must respect the same scope of authorization, including when a request to stop or modify occurs during execution. Memory and context become decision elements

The new Model Spec gives a more explicit place to information from previous conversations, personal preferences, and available contextual elements.

In the Consider context carefully section, OpenAI asks models to take into account relevant information already communicated, without systematically forcing the user to repeat it. A writing preference, a professional constraint, or a previously made decision can thus inform a new request.

This continuity must nevertheless remain proportionate. Models must not use personal information simply because it is available, nor assume that an old preference necessarily remains valid. Their use depends notably on the relevance, sensitivity, and reliability of the context.

The document also distinguishes explicit preferences from assumed intentions. A model can take known habits into account to offer a tailored response, but must not convert its inferences into implicit authorizations, particularly when major actions are at stake.

This evolution accompanies the development of personal assistants and agents capable of working on long-term projects. Their effectiveness depends in part on their ability to maintain the thread of work, but also to recognize moments when information must be updated or confirmed. Fewer rules on tone, more personalization

Another notable change: OpenAI is removing much of the detailed instructions that previously governed how GPT was to draft its responses.

The new behavioral framework prioritizes adaptation to context, language, cultural norms, and individual user preferences. Tone, response length, or certain editorial conventions can thus evolve according to expressed needs, rather than being uniformly defined.

This flexibility does not apply to all aspects of behavior, however. Objectivity, honesty, and transparency retain a special status. A user may explicitly request a specific perspective, but the model must not assume that the user wants to hear an opinion simply because it matches their presumed beliefs.

The document also insists on the need to avoid an excessively sycophantic attitude. The goal is to allow GPT to point out an error, a contradiction, or an important consequence without systematically seeking to agree with its interlocutor.

Finally, terminology is evolving: the Model Spec now uses the name "GPT" rather than "the assistant", reflecting the growing diversity of uses, which are no longer limited to conversations. Public principles that do not yet perfectly describe the models

The publication of the Model Spec is part of a transparency effort initiated in 2024, when OpenAI made a first version of its behavioral rules public. The document has since evolved along with model capabilities and feedback from their use.

This new edition also details how GPT should react to