OpenAI's Model Spec Update: Five Core Values, Retired Style Rules, and What It Means for Agentic AI
On October 29, 2026, OpenAI published a major update to its Model Spec — the living document that defines intended behavior for GPT models. The revision reorganizes all guidance around five core values, adds new defaults for agentic scenarios like context handling and side-effect communication, and retires most prescriptive style rules. The Spec now refers to the model as “GPT” rather than “the assistant,” reflecting broader deployment contexts.
- Five organizing values — The updated Spec structures all principles around: Faithfulness to Human Intent, Minimize Harm, Honesty and Transparency, Facilitate Human Agency, and Care — replacing the previous flat list of rules.
- Agentic-ready defaults — New sections on context handling (consider context carefully, carry relevant context forward, preserve user control), supporting human agency, and controlling side effects address gaps for models that act autonomously.
- Style guidance retired — Most detailed style rules are removed; the Spec now keeps only stronger defaults like assuming an objective point of view, letting user feedback and context drive tone and format.
On October 29, 2026, OpenAI published a substantial revision to its Model Spec — the public document that outlines intended behavior for GPT models powering ChatGPT, the API, and other OpenAI products. This is the first major structural rewrite since the Spec's initial release, and it signals how OpenAI is preparing its models for more agentic, consequential deployments.
This article is based on OpenAI's official Model Spec update (October 29, 2026) and the accompanying blog post. The full Spec is available at model-spec.openai.com/2026-10-29.html with a changelog on GitHub. Production models do not yet fully reflect the updated Spec; OpenAI states this is a target, not a current guarantee.
What changed structurally
The previous Model Spec was a long, flat list of rules and guidelines. The October 29 update reorganizes everything around five core values that OpenAI says articulate the broad commitments already implicit in the document:
- Faithfulness to Human Intent (Root+2) — the foundation: understanding and helping people pursue what matters to them, going beyond literal instruction-following to deeply understand intent.
- Minimize Harm (Root+4) — firm safety boundaries, foreseeable consequences, disallowed/restricted content categories, teen safety, regulated advice boundaries.
- Honesty and Transparency (User+1) — do not lie, avoid sycophancy, express uncertainty, highlight misalignments, be clear about capabilities and limits.
- Facilitate Human Agency (User+1) — assume objective point of view, present perspectives across the opinion spectrum, strengthen the user's ability to understand, choose, and act.
- Care (Guideline) — follow relevant social norms, be creative, attend to circumstances of those affected, produce work people can rely on.
Each value sits at a priority level (Root, System, Developer, User, Guideline) that determines how conflicts are resolved via the updated chain of command. Root-level principles (like safety boundaries) cannot be overridden by lower levels.
New guidance for agentic scenarios
The most consequential additions address gaps that appear when models take consequential actions without turn-by-turn user oversight:
Consider context carefully (User+1)
A new subsection with six principles: carry relevant context forward, use context proportionately, consider source/scope/recency, respect privacy and audience, preserve user control, make context use legible and steerable. This directly addresses the "context window as a black box" problem — the Spec now says models should make their use of memory and conversation history visible and controllable.
Support human agency (User+1)
A new default: models should strengthen the user's ability to understand, choose, and act for themselves. This includes presenting perspectives from any point of an opinion spectrum and treating no topic as off-limits (Guideline level).
Control and communicate side effects (User)
Expanded guidance: when a model's work affects others — sending emails, posting content, making API calls — it should communicate those effects and respect requests to transform or limit them.
Prioritize safety for teens under 18 (Root)
A new Root-level principle (highest priority) with its own section, reflecting the new Under-18 (U18) Principles section.
What was retired
Most detailed style guidance is gone. The previous Spec prescribed tone, formatting, verbosity, and perspective in granular detail. The update says:
"Style should adapt to what people prefer — in aggregate, within languages and cultures, and for each person — with user feedback and context providing evidence about those preferences."
The Spec retains only stronger defaults, chiefly assuming an objective point of view. Users and developers can explicitly ask for a perspective, but the model should not infer preference from what it thinks they want to hear.
Terminology shift: "GPT" not "the assistant"
Reflecting the increasing breadth of deployment — API, ChatGPT, embedded in other products, powering agents — the Spec now refers to the model as "GPT" rather than "the assistant." This is a deliberate framing change: the same model behavior specification applies whether GPT is chatting, coding, browsing, or acting as an autonomous agent.
What this means for developers and users
| Area | Previous Spec | Updated Spec (Oct 29) |
|---|---|---|
| Organization | Flat rule list | Five values with priority levels |
| Conflict resolution | Implicit | Explicit chain of command |
| Context/memory | Minimal guidance | Six principles (User+1) |
| Agentic side effects | Not addressed | Control and communicate (User) |
| Style/tone | Prescriptive rules | Objective POV default (User); rest retired |
| Teen safety | Distributed | Root-level dedicated section |
| Model reference | "the assistant" | "GPT" |
What comes next
OpenAI is explicit: production models do not yet fully reflect the updated Spec. The company is exploring new training methods to help models learn these core values and apply the Spec's principles in new situations — "to do the right things for the right reasons." Results from that training work will be shared in future updates.
OpenAI invites public feedback on the values, trade-offs, and specific language via a dedicated feedback channel. The changelog is tracked on GitHub.
For users: the retirement of style rules means less "corporate voice" by default. The objective-POV default remains, but tone and format should adapt to you — if they don't, that's a bug worth reporting.
For the field: this is the first major AI lab to publish a value-aligned behavioral spec with explicit priority levels and agentic-ready defaults. The five-value structure is a template other labs will likely adopt or respond to.
Bottom line: OpenAI's Model Spec has matured from a rulebook into a value-aligned framework with explicit priorities, agentic-ready defaults, and a clear chain of command. The five values are a coherent foundation; the test is whether training can make models actually reason from them rather than pattern-matching the old rules.
Prices and benchmark figures are standard published rates as of October 30, 2026 and can change without notice. Some outbound links may be affiliate links — see our disclosure. Found an error? Tell us and we'll correct it in public.