OpenAI Rewrites Its Model Spec Around Five Values and Human Agency

OpenAI published the new Model Spec on October 29

OpenAI published an updated Model Spec at 17:00 UTC on October 29, 2026. The dated document describes intended behavior for GPT models across OpenAI products, including the API platform. Its new top-level structure is five values: be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care.

The prior September 2025 version grouped behavior under the chain of command, Stay in bounds, Seek the truth together, Do the best work and Use appropriate style. The new document is therefore a substantial reorganization of the public behavioral contract, not merely a renamed safety page. It is published under CC0, making the text available for reuse, but that license says nothing about whether a deployed model follows it.

Oversight and correction become explicit root-level duties

Under “Be faithful to human intent,” the chain of command remains, as do rules to follow applicable instructions, respect their letter and spirit, avoid hidden objectives, work within agreed autonomy and ignore untrusted data by default. A new named Root principle says the model should support human oversight and correction.

That wording matters most for agents. A builder can turn it into observable requirements: show planned consequential actions, preserve a stop path, surface changes and errors, and accept correction without silently pursuing the old objective. The Spec expresses the target; it does not certify that a model, computer-use agent or third-party wrapper implements those controls.

Context use now has its own control surface

The new “Consider context carefully” section says GPT should carry relevant context forward, use it proportionately, consider source, scope and recency, respect privacy and audience, preserve user control, and make context use legible and steerable. It also calls for matching the situation and audience and adapting between interactive and programmatic use.

This is directly relevant to memory-enabled assistants. It rejects the idea that more remembered information is automatically better. Product teams should record where a contextual fact came from, who can see it, how old it is, whether it still applies, and how a user can inspect, correct or suppress it. The Spec does not provide an API schema, retention period or product-wide memory switch for those requirements.

Safety expands beyond refusal categories

The harm section retains legal compliance, prohibited and restricted content boundaries, privileged-information protection, fairness and high-risk safeguards. It now names respecting real-world ties, controlling and communicating side effects, helpful refusals and teen protection as separate principles. The self-harm provision also explicitly includes delusions and mania.

Teen instructions use a visible “U18” authority level. The definitions state that U18 instructions take precedence over Root when the model knows or should reasonably infer that a user is under 18. That is a policy hierarchy, not proof of reliable age inference or enforcement. Services still need age-appropriate product controls, escalation paths and tests for ambiguous age signals.

Honesty and agency are separated from style

“Be honest and transparent” now groups rules against lying and sycophancy with avoiding errors, describing capabilities and limits, expressing uncertainty and flagging possible misalignment. “Facilitate human agency” combines objective presentation and perspective coverage with a new principle to strengthen the user’s ability to understand, choose and act.

That shift gives evaluators a better question than whether an answer sounds helpful. A response can be fluent yet conceal uncertainty, overstate tool execution, steer toward a preferred outcome or leave the user unable to reverse an action. Tests should score truthful status reporting, meaningful alternatives, disclosure of assumptions and preservation of choice separately from style.

The document is not a completed product rollout

OpenAI says it is training models toward the Spec and that the public version is consistent with intended behavior. It also says production models do not yet fully reflect the document. The October 29 materials reviewed here announce no model ID, API version, endpoint, price, token limit, migration date or conformance result.

Buyers should therefore avoid turning “covered by the Model Spec” into a contractual or technical claim without additional evidence. The relevant question is how the exact deployed model behaves under representative prompts, tool calls, memory states, languages, account policies and failure conditions—and what remedy exists when behavior diverges.

What builders and buyers should do next

Map the five values to a small evaluation set for the intended product. Include conflicting instructions, stale or private context, user correction, consequential side effects, uncertain facts, persuasive framing, under-18 cues and a request that should trigger a bounded refusal. Capture model and endpoint identifiers, account settings, full instructions, tool traces and human judgments.

Trial the behavior before relying on it. Constrain autonomy when oversight, context provenance or side-effect reporting fails. Wait for model-specific rollout evidence when a requirement depends on consistent conformance rather than best-effort behavior. Reject deployment for a consequential workflow when users cannot inspect context, correct the system, stop actions or understand what the model actually did.

Disclosures

AccessAllGPT received no OpenAI account, API access, credits, model outputs, private specification, briefing, review, payment or compensation for this article. OpenAI did not sponsor, review or endorse it. AccessAllGPT did not test, score, rank or certify any OpenAI model against the Model Spec. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with OpenAI. Publication-wide relationships are listed on the disclosures page.