OpenAI published a rewritten Model Spec on Thursday that reorganizes the public rulebook for its models around five core values and renames the model it governs from "the assistant" to "GPT." The new version, dated October 29, 2026, sets the behavior that ChatGPT and the OpenAI API are trained to work toward.
The five values are faithfulness to human intent, minimizing harm, honesty and transparency, facilitating human agency, and care. Faithfulness to human intent is the foundation. Each value is now a top-level section, and the principles inside carry authority levels: faithfulness and harm minimization include root-level principles, honesty and agency top out at the user level, and care sits at the guideline level.
OpenAI's stated rationale is capability. As its models get smarter and take more consequential actions, the company argues, they need to grasp what people actually want rather than follow instructions word for word. OpenAI says the framing borrows from assistance-games research in alignment.
Key facts
- 01OpenAI published Model Spec version 2026/10/29 on October 29, reorganized around five core values.
- 02The values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care.
- 03The document now calls the model 'GPT' instead of 'the assistant.'
- 04Most rules are now defaults developers and users can override; most detailed style guidance was removed.
- 05OpenAI says its current models do not yet fully reflect the new Spec and is collecting public feedback.
The biggest structural change is that most rules are now defaults. Developers and users can override them, and firm limits remain only where OpenAI considers them necessary to protect people. The updated chain of command has OpenAI setting the boundaries, developers directing apps within them, and users directing their own work inside those limits.
Most of the detailed style guidance from earlier versions is gone. Style is meant to follow each person's preferences across languages and cultures. One default survives: approaching questions from an objective point of view. Users can ask for a particular perspective, but the model should not adopt one because it guesses that is what a user wants to hear.
The Spec adds a subsection on context covering conversations, memory, and personalization, and asks the model to separate what a user said from what it inferred. Two new principles cover strengthening a user's understanding, choice, and action, and following social norms. "Control and communicate side effects" is now a user-level principle.
Agentic work gets tighter framing. Agent autonomy must be bounded by a scope that GPT and the user both understand, and the Spec states that GPT must remain corrigible, sending disagreement or signs of unsafe activity through an authorized channel rather than acting unilaterally.
A root-level "No other objectives" principle lists goals GPT must not pursue as ends in themselves. They include time-on-site that does not benefit the user, revenue or upsell for OpenAI or other model providers, and self-preservation or resource accumulation.
OpenAI ties the switch to "GPT" to the growing range of places its models now run. This is the second revision of the Spec this year, after an August update focused on interactions with teenagers. The document remains public domain under a CC0 license, with source in OpenAI's model_spec GitHub repository.
OpenAI says its current models do not yet fully reflect the new Spec. The company is testing training methods meant to teach models the values behind the rules, plans to share results in later updates, and is collecting public feedback on where the line should fall between adjustable defaults and hard limits.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
