OpenAI rebuilds Model Spec around five values, drops most style rules

Summary

• OpenAI reorganized its public Model Spec around five core values governing how its models should behave.
• The document now adds guidance on context, oversight and agency, while retiring most detailed style instructions.
• OpenAI said the Spec describes intended behavior that its models do not yet fully reflect.

The latest

OpenAI has rewritten its Model Spec around five underlying values — faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care — in what the company described as an attempt to make the reasoning behind its rules explicit. The update also adjusts some behavioral defaults and retires much of the document's detailed style guidance. OpenAI said the Spec sets out behavior it is working toward, not behavior its models already achieve.

Details

• The core framing: OpenAI said the five values articulate commitments already spread across the Spec: serving human purposes, being truthful, minimizing harm, respecting people's freedom to form their own views, and taking their circumstances seriously. Grouping them, the company said, explains why the Spec asks models to behave as it does and exposes where guidance was thin.
• Faithfulness defined: The company called faithfulness to human intent the foundation of the approach — going beyond literal instruction-following to understand what OpenAI, developers and users intend, and acting when that understanding is incomplete. It also covers helping people reflect and change their views without quietly steering them or overriding their choices.
• The chain of command: OpenAI set out a layered authority structure: the company writes the Spec, sets boundaries and trains models to respect them; developers direct applications within those boundaries; and users direct the model's work subject to those constraints. The new faithfulness section incorporates an updated chain of command explaining how conflicting instructions are resolved.
• The other four values: Harm minimization covers firm safety boundaries and foreseeable consequences of a model's actions. Honesty and transparency require models to be truthful about what they know, do and are uncertain about. Human agency means strengthening people's ability to choose and act. Care means following relevant social and professional norms and producing reliable work.
• New sections: A subsection titled consider context carefully explains how conversations, memories and other contextual evidence should shape a model's reading of user intentions. OpenAI also added defaults on supporting human agency and following relevant social norms, and expanded existing guidance on controlling and communicating side effects.
• Style guidance retired: Most detailed style rules were dropped, with OpenAI arguing style should adapt to what people prefer in aggregate, within languages and cultures, and individually. Stronger defaults remain, including approaching questions objectively; users and developers can request a particular perspective, but the model should not infer one from what it thinks they want to hear.
• The naming change: Citing the widening range of use cases and deployment settings, OpenAI said it now refers to the model as "GPT" rather than the assistant throughout the Spec.
• Training work: OpenAI said it has been exploring new training methods to help models learn the core values and follow the Spec's behavioral principles, aiming to build an understanding of why those principles matter and how to apply them in unfamiliar situations. The company did not detail the methods, saying results and lessons would come in future updates.
• Worked example: Illustrating faithfulness, OpenAI described a model helping someone compare job offers by drawing on priorities the user previously shared, such as predictable hours, while making those assumptions visible and easy to correct.
• Open feedback: The company invited readers to say which parts of the Spec they find compelling, where they disagree with the balance it strikes, and what it overlooks, saying feedback on values, trade-offs and specific language would inform revisions. No deadline or review schedule was attached.

Background

The Model Spec is OpenAI's public description of how it wants its models to behave. The company said the document is refined through research, experience deploying models, public feedback on the Spec and model behavior, and discussions with users and outside experts.

Between the lines

Dropping most style rules while keeping firm safety boundaries draws a sharper line between what the company treats as negotiable and what it does not: presentation shifts toward user preference, while harm limits stay fixed. The layered chain of command — OpenAI, then developers, then users — makes that hierarchy explicit rather than implied. OpenAI's acknowledgment that models do not yet fully reflect the Spec frames the document as a target, leaving the gap between stated intent and deployed behavior as the measurable test.

What's next

Watch for OpenAI's promised publication of results and lessons from its new value-training work, and for revisions to the Spec drawn from the public feedback the company solicited on its values and trade-offs.

Source: OpenAI