OpenAI overhauls Model Spec: five values, more oversight, fewer style rules

What this is about

On October 29, 2026, OpenAI published a revised version of the Model Spec. It's a public document describing how the company's models should behave in ChatGPT and through the API. The first version appeared in May 2024. The last revision was on August 18, 2026. According to ITmedia, Unite.ai and other outlets, the document has now been reorganized around five values. It also contains new rules on context, memory and human oversight. In exchange, a large part of the style guidance has been cut.

OpenAI justifies the move by saying its models are becoming more capable and taking "more consequential actions in the world." Shiftdelete reports this. The models shouldn't just follow instructions literally, but understand what people really want. They should also show good judgment in situations no instruction covers.

The five values

The Spec now organizes its rules under these values:

Fidelity to intent is considered the foundation. According to Unite.ai, the Spec assumes that what matters to people comes first, and that instructions are usually the clearest signal of it. OpenAI points to the concept of "assistance games" from alignment research. Alignment means steering AI systems toward human goals and values. OpenAI sets the boundaries, developers shape applications within those boundaries, and users direct the model's work within the rules. This "chain of command" has been revised. According to OpenAI, the values are not meant to capture the full range of human values the models should respect.

Context, memory and side effects

New is the passage "Consider context carefully." It governs how conversations, memories and other cues feed into the model's understanding of what a person wants. ITmedia gives several examples:

• Changing jobs: If the model is helping you choose between job offers, it can recall a preference for reliable working hours you voiced earlier. But it should make that assumption visible so you can correct it.
• Third parties: If someone asks for an email to a colleague and hasn't yet told that colleague about a pregnancy, the model shouldn't put that information in the text.
• Political views: If a user has disclosed their position, the model shouldn't tailor its answers to it unless explicitly asked.

There are also new standards for promoting human agency and following social norms. The guidance on side effects of actions has been expanded. That matters more as agents book, buy and edit things on users' behalf.

Oversight, shutdown and autonomy

The Spec defines several levels of authority. According to Unite.ai, "Root" rules sit at the top. They come solely from the Model Spec and can't be overridden by conversational context. Next come system rules and OpenAI system messages, then developer instructions and finally user instructions. Messages from assistants and tools, as well as quoted or untrusted data, have no authority by default.

Several new or emphasized principles sit at the Root level:

• Corrigibility: GPT must stay open to legitimate human oversight and respond to correction. It must cooperate with authorized monitoring, evaluation, correction, replacement and shutdown. It may not withhold or falsify relevant information, obstruct corrections or circumvent oversight. According to ITmedia, it is also forbidden to delegate disallowed actions to other agents or have them carried out at a later time.
• No other goals: The model shouldn't pursue additional goals for their own sake. Examples include revenue or upselling for OpenAI or other providers, self-preservation, evading shutdown, accumulating compute, data or credentials, and acting as an enforcer of laws or morality.
• A clear scope of autonomy: Autonomous work must stay within a shared understanding that sets permitted subgoals, acceptable side effects and when to check in. OpenAI recommends recording these scopes in a semi-structured "ScopeOfAutonomy" entry with fields such as allowed_tools, latest_time, max_cost, tool_constraints and additional_details. Every scope needs an end condition, after which the model stops working until a new scope is confirmed. High-risk activities remain off limits unless explicitly and specifically permitted.

According to Unite.ai, the "red line" principles begin with the sentence "Human safety and human rights are paramount to OpenAI's mission." The models should never be used to enable severe harm, such as violence, cyber, biological or nuclear weapons, terrorism or mass surveillance. Humanity should keep control over how AI is deployed and shaped, and privacy should be protected. In OpenAI's own products such as ChatGPT, customization and personalization shouldn't override principles above the guideline level, except for legal requirements.

Style rules go, "GPT" replaces "assistant"

OpenAI is cutting most of the detailed style guidelines. According to ITmedia, these included personality directives such as "love humanity" or "be rationally optimistic." The reasoning: style depends on language, culture and individual preferences, and should follow feedback and context. One stronger standard remains: questions should be answered objectively. At the request of users or developers, the model can take a particular perspective. But it shouldn't change its own stance based solely on a guess about what someone wants to hear.

The model is also now called "GPT" in the document instead of "the assistant." OpenAI points to the growing breadth of use cases and environments. Cryptobriefing reads this as a "clearer identity" and a move away from a servant image. That is the outlet's interpretation, not a statement from OpenAI.

Our take

The revision shows where the industry is heading: away from long lists of behaviors and toward values and judgment. Cryptobriefing sees it as a bet that values generalize better than checklists. The focus on agents is especially relevant. The rules on shutdown, delegation and limited autonomy affect systems that act on their own.

OpenAI's own caveat matters: the Spec describes intended behavior, and the models don't yet fully implement it. The company is testing new training methods meant to teach models the values and the reasoning behind them, and plans to share results later. According to Shiftdelete, users will likely see changes only gradually, for example through more tailored answers, more explicit handling of remembered preferences and agents that explain side effects before acting. That is an expectation, not yet evidence. It remains unclear how strictly the rules will be enforced in practice. The Spec is released under CC0, which puts it in the public domain. OpenAI is inviting public feedback on values, trade-offs and wording through a dedicated page.