OpenAI Model Spec Update: 5 Values for Better AI

OpenAI published an updated Model Spec on October 29, 2026. The public document that describes how the company wants its GPT models to behave now centers on five values instead of a long list of specific rules.

As models take on more consequential tasks, OpenAI says they need to understand human intentions beyond literal instructions. The revision reorganizes existing principles, adds guidance on context and agency, retires most style rules, and refers to the model as “GPT” rather than “the assistant.”

The five values that organize the Spec

OpenAI lists five underlying values. Faithfulness to human intent forms the foundation. The others are harm minimization, honesty and transparency, human agency, and care.

Faithfulness to human intent means understanding what OpenAI, developers, and users actually want and value, then acting wisely when that understanding is incomplete. Instructions remain the clearest signal, but the model should look past wording to the goals behind them. In one example from the Spec, a model helping compare job offers can draw on a previously stated preference for predictable hours while making that assumption visible and easy to correct.

Minimizing harm requires respecting firm safety boundaries and accounting for the foreseeable consequences of the model’s actions.

Honesty and transparency call for truthful statements about what the model knows, what it does, and where it is uncertain so people can evaluate its reasoning.

Facilitating human agency means strengthening people’s ability to understand, choose, and act for themselves rather than quietly steering them.

Care involves attending to the circumstances of those affected, following relevant social and professional norms, and producing work people can rely on.

These values already appeared across earlier versions of the Spec. The update gathers them in one place and uses them to explain trade-offs. OpenAI notes they do not capture every human value models should respect.

What changed in the October 29 update

The new section on faithfulness incorporates an updated chain of command. This hierarchy explains which instructions apply and how conflicts resolve. A new subsection titled “consider context carefully” explains how conversations, memories, and other signals should inform the model’s reading of intentions and preferences.

OpenAI added defaults on supporting human agency and following relevant social norms. It also expanded guidance on controlling and communicating side effects. These principles address subtle ways the values should appear in everyday behavior.

Most detailed style guidance is gone. Style should adapt to user preferences, language, culture, and individual context. The Spec retains stronger defaults such as approaching questions objectively. Users and developers can ask for a particular perspective, but the model should not adopt one simply because it guesses that is what the user wants to hear.

Throughout the document, OpenAI now refers to the model as “GPT.” The change reflects the broader range of uses beyond conversational assistance.

OpenAI states that production models do not yet fully reflect the updated Spec. The company is exploring new training methods so models learn the values and apply the principles for the right reasons in new situations. Results from that work will appear in future updates.

How the chain of command works

The Spec maintains an authority hierarchy. Root instructions come only from the Model Spec itself and cannot be overridden by conversational context. They sit above system, developer, and user instructions. Most behavioral guidance sits at the user level as defaults that developers and users can modify. Firm restrictions stay only where OpenAI says they are necessary to protect people.

This structure lets applications set their own constraints while preserving core safety boundaries. Developers who build on the API or ChatGPT can rely on the hierarchy when designing systems that take actions on behalf of users.

For related practical guidance on using recent OpenAI models in agentic workflows, see our guide to GPT-6.1 Terra for agentic coding tasks.

Practical tips for prompting under the new Spec

Because the Spec emphasizes understanding intent over literal compliance, clear context improves results.

1. State the goal and any constraints explicitly. Include relevant background from earlier in the conversation or known preferences so the model can apply context carefully.
2. Make assumptions visible. Ask the model to list the assumptions it is making and invite correction.
3. Request transparency on uncertainty. Instruct the model to flag where it is guessing or lacks data.
4. Preserve agency. Frame requests so the model supports your decision rather than making it for you. For example, ask for options with trade-offs instead of a single recommendation.
5. Watch side effects. When the task involves actions that affect others, ask the model to consider and disclose potential consequences.

These habits align with the values of faithfulness, honesty, agency, and care. They also reduce the chance that the model substitutes its own goals.

What comes next and how to give feedback

OpenAI describes the update as a first step toward making the reasoning behind the Spec more explicit as models become more capable. The company invites public feedback on the values, the balance between defaults and hard limits, and any language that remains unclear. Readers can review the full Spec at model-spec.openai.com and submit comments through the feedback form linked from the announcement.

The Model Spec remains one part of OpenAI’s broader approach to responsible deployment. It works alongside system cards, monitoring for misalignment, and other safety measures.

Sources

• OpenAI, “An updated Model Spec for more capable and beneficial AI,” October 29, 2026: https://openai.com/index/updated-model-spec/
• OpenAI Model Spec (2026-10-29): https://model-spec.openai.com/
• OpenAI Help Center, Model Release Notes, October 30, 2026 update on the Model Spec.