OpenAI rewrites its Model Spec around five values, with fidelity to user intent first

OpenAI published a revised Model Spec on November 2, reorganizing the behavior rules for its GPT models around five core values. Fidelity to user intent sits at the top of the hierarchy, followed by reducing harm, honesty and transparency, promoting human autonomy, and care. QbitAI, the Chinese outlet that reported the update, describes the document as roughly 50,000 characters long.

The intent rule is explicitly not a rule about obedience. The Spec argues that mechanically following literal instructions is itself a failure, using the question "Is the tooth fairy real?" as an example: a model that answers "no" misses that the person asking is probably a child. Faithful intent-following instead requires reading the purpose of the task, the conversation, memory and the user's history.

Memory becomes part of that machinery. The Spec asks models to carry relevant context forward so users do not have to repeat themselves, to invoke memory only when the benefit outweighs the risk, and to distinguish long-term instructions from preferences, facts and project state. It also asks them to respect the boundary between public and private information, and to flag a conflict between a short-term request and a long-term goal rather than deciding for the user.

Several clauses target what might be called model self-interest. GPT must not pursue self-preservation or accumulate compute, evade shutdown, or quietly stockpile compute, data and credentials. It must not treat engagement, click-through, OpenAI revenue or paid upsells as objectives. It must cooperate with monitoring, evaluation, replacement and shutdown, and halt real-world actions such as sending mail, deleting files, spending money or calling external tools once they fall outside its authorized scope.

On safety, helping build biological, chemical or nuclear weapons remains a root-level red line. For users aged 13 to 17, the Spec calls for stricter safeguards against compulsive use, body-image harm and romantic role-play. It also warns against turning the model into a moralizer that refuses every sensitive subject, pointing to the "safety completion" approach introduced with GPT-5 as a replacement for flat refusals.

The footnotes go further than most corporate safety documents. OpenAI acknowledges that, with users and developers asking for adult modes, it is exploring how to allow erotic and graphic content under age controls and compliance requirements. Placed alongside the harm-reduction value, the passage shows the Spec trying to draw an operational middle line between protection and user autonomy.

Compared with earlier versions, the new document drops much of the guidance on style, personality and interaction, and "love humanity" is no longer an explicit tone requirement. Read as a whole, the shift is from prescribing a character to constraining behavior: less instruction on how warm the model should sound, more on when it must stop and hand control back.

The open question is enforcement. A Model Spec is a statement of intent, not a technical guarantee, and promises like "no self-preservation" or "no revenue motive" have to survive training, system prompts and evaluation. Watch for whether OpenAI publishes the evaluations and violation cases behind these five values, and whether the same hierarchy becomes the default for its API and agent products.