OpenAI’s Model Spec drops most style rules, adds five values

OpenAI on October 29 rewrote its Model Spec, the public document that describes how its models should behave, around five core values and retired most of its detailed style guidance. Tone and format are now meant to follow what people prefer, using user feedback and context as the evidence. Approaching questions objectively stays a default: GPT should take a particular perspective when a user or developer explicitly asks for one, and should not infer that preference from what it thinks they want to hear. OpenAI’s own caveat is that the Spec describes behavior it is working toward, and its models do not yet fully reflect it.

The five values, as OpenAI’s announcement lists them, are faithfulness to human intent, minimizing harm, honesty and transparency, facilitating human agency, and care. Faithfulness is the foundation. It means going beyond literal instructions to what OpenAI, developers and users intend and value, and acting wisely when that understanding is incomplete. OpenAI writes the Spec and sets boundaries, developers direct applications within them, and users direct the model’s work within those constraints. Firm boundaries stay where OpenAI says they are needed to protect people and society. The rest mostly takes the form of defaults that users and developers can override.

What the new sections add

The new section on faithfulness starts from the idea that what people value comes first, with instructions ordinarily the clearest signal of it. OpenAI says it borrows this approach from the assistance-games framing of alignment. The section folds in an updated chain of command, the part of the Spec that decides which instructions apply and how conflicts between them are resolved.

A new subsection, “Consider context carefully,” covers how conversations, memories and other contextual evidence should shape the model’s read of what people intend and prefer. The changelog for this release, v2026.10.29, says it covers using context, memory and personalization in relevant, proportionate ways while respecting privacy and user control. The same release adds explicit guidance on supporting human oversight and correction. OpenAI gives the example of comparing job offers: the model might draw on priorities the person shared before, such as predictable hours, while making those assumptions visible and easy to correct.

The update also adds defaults on supporting human agency and following relevant social norms, and expands the guidance on controlling and communicating side effects. That guidance first appeared on September 12, 2025, as one of the Chain of Command sections in a release that added principles for agents that take actions in the world. This release makes it a user-level principle. In the Spec’s chain of command, user is the fourth of five authority levels, after root, system and developer.

Style moves to feedback

Most of the detailed style guidance is gone. OpenAI says style should adapt to what people prefer in aggregate, within languages and cultures, and for each person, with feedback and context as the evidence. The Spec still keeps some stronger defaults, and objectivity is the one OpenAI names.

OpenAI has run into trouble with feedback before. A GPT-4o update released on April 25, 2025, was rolled back four days later, and OpenAI said the update “focused too much on short-term feedback and did not fully account for how users’ interactions with ChatGPT evolve over time,” according to a Georgetown Law tech brief. The Spec’s December 18, 2025 revision also said that time-on-site and clicks should count only insofar as they align with genuine user value.

Text first, training later

The first draft, published May 8, 2024, was written as guidelines for researchers and data labelers producing RLHF training data. At the time, OpenAI said it was working on techniques that would let models learn directly from the Spec. OpenAI now says it has been exploring new training methods to teach models the five values and the reasons behind the Spec’s principles, “to do the right things for the right reasons.” Results from that work are still only promised for future updates.

The rewrite also renames the model. The Spec now calls it “GPT” instead of “the assistant,” a change OpenAI ties to the growing breadth of its models’ use cases and deployment settings. OpenAI is asking readers to tell it “where you disagree with the balance it strikes, and what it overlooks or leaves unclear.”