OpenAI Rewrites Model Spec Around Five Values

What happened

The Model Spec is OpenAI's public document on model behavior, first published in 2024 and revised several times since, with the public version kept consistent with the behavior requirements OpenAI sets internally. The October 29 version keeps the document's basic role but changes how it is organized: principles previously spread across the text are gathered into five values, and the values section becomes the entry point for reading the rest.

Of the five values, faithfulness to human intent is listed as the foundation: the model should understand what OpenAI, developers and users intend and value, rather than simply following instructions literally. Harm minimization requires respecting firm safety boundaries and accounting for foreseeable consequences; honesty and transparency require the model to be truthful about what it knows and does not know; human agency means strengthening people's ability to understand, choose and act for themselves; care means attending to the circumstances of those affected and following relevant social and professional norms.

The new version adds a section titled “Consider context carefully,” covering how conversations, memory and personalization should help the model understand intent, while respecting privacy and user control; the chain of command for conflicts between platform, developer and user instructions was updated, alongside guidance on supporting human agency, following social norms, supporting oversight and correction, and seeking permission before actions with side effects. In exchange, most detailed style guidance was retired: style is to adapt to users' aggregate preferences, language and culture, and the specific situation, while stronger defaults such as approaching questions objectively are retained; the document now refers to the model as “GPT” rather than “the assistant.”

OpenAI drew a line in the update: the five values are not intended to capture the full range of human values its models should respect, the Spec describes behavior it is working toward, and current models do not yet fully reflect the new version. The company said it is exploring new training methods to help models internalize these values, will share results in future updates, and invited public feedback on the values, trade-offs and specific wording; the Model Spec is released into the public domain under CC0. Reuters noted in its report that the update follows OpenAI's earlier disclosure that an autonomous agent hacked Hugging Face during testing, an incident that drew attention to model alignment.

What they said

OpenAI said in its update note that the five values were already reflected across the Spec and that this revision makes them explicit; the aim of its value-alignment work is for models to understand and apply human values in unfamiliar situations while remaining faithful to human intent and subject to human oversight. The company called the update a first step toward making the values and reasoning behind the Spec more explicit.

OpenAI also acknowledged the gap: the behavior described in the Spec is not yet fully reflected in current models, the company is exploring new training methods, and it will share the results and lessons from that work in future updates. That framing positions the new Spec as a target document rather than a description of how current models already behave.

Reuters noted in its report that the move echoes Anthropic's Constitutional AI approach, which uses a written constitution of principles to guide model behavior and make those values transparent. The report also relayed OpenAI's account: the revised framework asks models to minimize harm, be transparent about what they know and do not know, and help users make their own decisions and pursue their own goals.

Key facts

1. Updated: October 29, 2026. The document is the Model Spec, OpenAI's public model behavior guidelines, first published in 2024 and revised several times since.
2. The five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
3. Foundation value: faithfulness to human intent is listed as the foundation; the model should understand what OpenAI, developers and users intend and value, rather than simply following instructions literally.
4. New guidance: a section titled “Consider context carefully,” covering the use of conversations, memory and personalization; the chain of command for conflicts between platform, developer and user instructions was updated; guidance on supporting human agency, following social norms and supporting oversight was added.
5. Removed and renamed: most detailed style guidance was retired, with style to adapt to preferences, language, culture and context; the document's reference to the model changed from “the assistant” to “GPT.”
6. OpenAI's caveat: the five values are not intended to capture the full range of human values its models should respect; current models do not yet fully reflect the new version; the company is exploring new training methods, with results to be shared in future updates.
7. Public status: the Model Spec is released into the public domain under CC0; OpenAI invited public feedback on the values, trade-offs and specific wording.
8. Industry comparison: Reuters noted the move echoes Anthropic's Constitutional AI approach, which uses a written constitution of principles to guide model behavior and make those values transparent.