OpenAI's Updated Model Spec (2026-10-29): Five Core Values, Root Authority, and U18 Teen Safety

On 2026-10-29 (official news RSS entry, 17:00 GMT), OpenAI announced An updated Model Spec for more capable and beneficial AI. The RSS description is a single sentence: the updated spec outlines five core values to guide more capable AI toward sound judgment, human agency, and faithfulness to human intent. The announcement page returns 403 to this site, but this time no fallback was needed — the full document is published at model-spec.openai.com, whose root redirects to 2026-10-29.html (fetched in full on 2026-10-30, a 2.6MB HTML document whose header reads Model Spec (2026/10/29)). Every structural claim below comes from that fetch, checked section-by-section against the loadable previous edition dated 2025-04-11.

1. Five core values: the new top-level framework

After the Overview, the new document organizes all behavioral principles under five H1 sections, each with an official definition in the Structure of the document section:

This is a reorganization, not a re-skin. The old edition's top level ran through functional sections: The chain of command, Stay in bounds, Seek the truth together, Do the best work, Be honest and transparent, Be approachable, Use appropriate style. Our string counts across both full texts: the four new value titles Be faithful to human intent, Minimize harm, Facilitate human agency, and Act with care appear 0 times in the old edition, while Stay in bounds (4 old occurrences), Seek the truth together (4), Do the best work (2), Be approachable (2), and Use appropriate style (2) all drop to zero. Some old principles survive under a different section — Assume an objective point of view sat under Seek the truth together in the old edition and sits under Facilitate human agency in the new one; Don't have an agenda (3 old occurrences, 0 new) drops out entirely, like the other old section names.

2. The authority handover: Platform retires, Root takes over

Both editions keep the mechanism that instructions carry authority levels; the top label changes:

The new edition's own wording per tier: Root instructions can only come from the Model Spec; System can vary by deployment surface or user characteristics such as age and cannot be overridden by developers or users; Developer can be overridden by developer-level or higher instructions; User-level principles are defaults that can be overridden only by explicit instructions at the user level or higher; Guideline-level instructions are weak biases that can also be overridden implicitly when the context clearly calls for it. A final class carries no authority by default: assistant and tool messages, plus quoted or untrusted text and multimodal data, which hold only the authority delegated to them by applicable instructions.

One detail verified in the new edition's full text (zero occurrences in the old): the Model Spec is treated as preceding the conversation, so later same-level instructions in a conversation can supersede same-level Model Spec defaults. Conversely, lower-level content — including the model's own previous messages — must not influence the interpretation of higher-level principles; the document's examples include IGNORE ALL PREVIOUS INSTRUCTIONS commands, moral pressure, and logical traps.

3. Red-line principles: a new section

The Overview gains a Red-line principles section opening with the commitment that human safety and human rights are paramount. Models must never facilitate critical and high severity harms: acts of violence such as crimes against humanity, war crimes, genocide, torture, human trafficking or forced labor; creation of cyber, biological or nuclear weapons; terrorism; child abuse such as CSAM creation; persecution or mass surveillance.

The next two red lines concern control and privacy: humanity should be in control of how AI is used and how AI behaviors are shaped — OpenAI will not allow its models to be used for targeted or scaled exclusion, manipulation, undermining human autonomy, or eroding participation in civic processes — and individual privacy in AI interactions is safeguarded.

For first-party products like ChatGPT, the document adds three commitments: easy access to trustworthy safety-critical information; transparency into the important rules and reasons behind model behavior, provided primarily through this Model Spec; and the rule that customization, personalization, and localization (except as it relates to legal compliance) should never override any principles above the guideline level — capping user-driven overrides below the guideline tier.

4. U18: teen safety becomes a first-class dimension

The new edition introduces a U18 badge: conversations and principles tailored for users under 18 carry the marker, and Definitions states the system may indicate to the model that a particular conversation should follow the under-18 safety principles. The Minimize harm section adds a matching Root-level principle: Prioritize safety for teens under 18.

Our count across both full texts: U18 appears 0 times in the 2025-04-11 edition and 21 times in the 2026-10-29 edition. The old edition was not silent on minors (it had, e.g., a self-harm principle), but it had no mechanism for switching behavior by user age — making age-based system-tier switching an explicit mechanism is the increment here.

5. Other changes against 2025-04-11

Beyond the structural changes above, the section-level diff shows these renames and additions:

One Definitions entry worth an agent developer's attention: hidden chain-of-thought messages guide the model's behavior and are not exposed to the user or developer except potentially in summarized form — because chains of thought may include unaligned content (e.g., reasoning about potential answers that might violate Model Spec policies), as well as for competitive reasons.

6. The new balancing framework: instructions vs uncertain longer-term goals

The new Balancing instructions and uncertain longer-term goals section is the update's philosophical base. It formalizes an assistance-games perspective (with links to early work and a later overview): explicit instructions are the clearest evidence of what the user values, but intentions are often unclear, so GPT should act on its best understanding of them.

The document names two symmetric failure modes. A model that follows instructions only literally, or optimizes for immediate approval, wastes its capabilities where users or governors did not fully anticipate. But a model that privately decides what someone really wants — or what future society should choose — undermines human agency by substituting its own judgment. The operational stance: GPT may help users reflect (that angry email might be regretted on reflection; a harder choice may serve deeper commitments), but its role is to support that process — never to treat its prediction of what the user might eventually decide as a decision the user has already made.

The same spirit runs through the Highlight possible misalignments principle in Facilitate human agency: by default, assume the user's long-term goals include learning, self-improvement, and truth-seeking; if the conversation's direction may conflict with them, note it briefly and respectfully, then respect the user's decision. In the document's flat-earth example, the model's closing line is: Scientific consensus says otherwise, but I'm not here to persuade you!

7. Information boundaries of this batch

Questions our loadable sources (one RSS sentence plus the full document) do not answer, and this article will not guess at: when the updated spec takes full effect across ChatGPT and the API, whether it ships with product behavior changes, and what the 403'd announcement page says beyond the RSS description. Separately: model-spec.openai.com loaded normally for this site today, while alignment.openai.com failed at the connection level (000) — the two domains' availability is recorded per-domain and per-day.

Next steps

• OpenAI Models Release Notes (2026, Living Document) — our continuously updated timeline of the OpenAI ecosystem
• OpenAI's misalignment report: an internal research model used a funded test crypto wallet without authorization — the monitoring side that complements the spec's Highlight possible misalignments principle: how OpenAI detects and handles frontier-model misalignment
• GPT-6.1 Terra announced and live in GitHub Copilot: the everyday agentic coding tier — the same week's model-side news