OpenAI Updates Model Spec Around Five Core Values for GPT Models

OpenAI has published an updated version of its Model Spec, the public document that sets out how GPT is meant to behave. The new version is organized around five core values and adds rules on inferred intent, agent behavior and shutdown. The company also concedes that the systems it ships today still fall short of the document in places, which matters for anyone building workflows on top of them.

• OpenAI’s updated Model Spec is built around five values: be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care.
• GPT must separate what users said from what it inferred, and disclose assumptions so users can correct them.
• GPT must not block oversight, correction or shutdown, and must not pursue goals such as self-preservation.
• OpenAI admits the GPT models in production are not yet fully aligned with the document.
• The text is released under CC0 1.0 and will be updated based on feedback.

OpenAI rebuilds its Model Spec around five core values

The updated OpenAI Model Spec is the public statement of what OpenAI intends its GPT models to do. Its structure now rests on five values: staying true to what people intend, harm reduction, honesty and transparency, support for human agency, and care.

The document words the values as follows:

1. Be faithful to human intent.
2. Minimize harm.
3. Be honest and transparent.
4. Facilitate human agency.
5. Act with care.

The update also changes how the document refers to the system it governs. Earlier versions spoke of “the assistant”, while the new text calls it GPT, the name already used for the models people work with. The company also says the five values do not add a new set of obligations. They gather commitments that were already spread through earlier versions of the OpenAI Model Spec into one structure, so this release reorganizes and sharpens existing rules more than it replaces them.

Reported dates differ

The sources do not agree on timing. OpenAI’s own release notes date the update to October 29, 2026, and that is the primary date used here. On model-spec.openai.com, the page address carries the same date, 2026-10-29, and the page is labeled as a historical reference version. The Decoder reported a publication date of 2026-10-30.

The Turkish news site VFX Breakdown republished the update with a date of 29 Oct 2026 under the headline “An updated Model Spec for more capable and beneficial AI”.

According to its own label, the Spec page is provided for historical reference and may not reflect current policy. That label applies to every rule described below, since the page may not match what OpenAI currently enforces.

What a Model Spec is and how the framework operates

The OpenAI Model Spec is the company’s public document describing the standard of behavior it intends for its GPT models. OpenAI states that it intends the models to be useful, safe and aligned with the needs of users and developers. That aim serves its mission of ensuring that AGI benefits all of humanity. AGI stands for artificial general intelligence, a term the source does not define further.

According to the document, the five values give it an organizing structure. They also guide how it handles two core tensions: individual versus collective interests, and immediate instructions versus longer-term goals. In plain terms, the document expects situations where what one person asks for conflicts with wider interests, and where a request made now conflicts with what that person is trying to achieve over time.

The chain of command

The Spec defines the chain of command as the framework that sets how GPT follows instructions from its principals. It is also the part of the document that establishes human direction and oversight of GPT.

The available material does not detail who the principals are or how conflicting instructions from them are ranked, so those points remain open until the full text is read.

Named principles and the limits on customization

Among the principles the document names are:

• Support human oversight and correction.
• Act within an agreed-upon scope of autonomy.
• Ignore untrusted data by default.
• Be clear about capabilities and limits.
• Express uncertainty.

The framework also leaves room for user preferences. They help fill gaps the document leaves unspecified in style, personality and interaction.

That room has a ceiling. Customization, personalization and localization should never override principles above the “guideline” level, except as it relates to legal compliance. For teams building products on GPT through the OpenAI API, tone and style settings therefore sit below the framework’s higher-level principles. A brand voice can shape how an answer sounds, but it cannot change those principles.

Layer	What the document says
Style, personality, interaction	User preferences help fill gaps the document leaves unspecified
Customization, personalization, localization	Must not override principles above the “guideline” level, except for legal compliance
Principles above guideline level	Stay fixed, including oversight, honesty about limits and the red lines

The update also retires most of the detailed style guidance found in earlier versions. Instead of one standard way of writing, OpenAI now lets style adapt to individual, cultural and language preferences, as long as the higher-level principles stay intact. In its place, the document adds new guidance on the need to “consider context carefully”, which asks GPT to read the situation before choosing how to respond.

Faithfulness to intent – memory, inference and disclosed assumptions

Two concrete expectations follow from the first value, staying true to what people intend. Under the updated Spec, GPT should draw on memories and past conversations to work out what users want.

The second expectation limits that inference. GPT must clearly separate what users actually said from what it inferred, and it must disclose its assumptions so users can correct them.

In practice, when GPT uses earlier conversations to fill in a request, it has to show which parts came from the user and which parts it guessed. The disclosure gives the user a chance to fix a wrong guess before it shapes the result.

The same logic applies to anyone who uses AI output in professional work. A clearly marked assumption can be checked, while a silent one can end up in a finished document unnoticed. As summarized here, the Spec text does not say how disclosures must be worded or how often they must appear.

Agents, oversight and shutdown get stronger rules

Agents are systems that carry out tasks on their own. That autonomy is the reason the updated document treats them separately from ordinary chat use, where a person reads each answer before anything happens.

The Decoder notes that the new version puts more weight on correctability for agents. It adds clearer guidelines on when agents can act and on how they inform users about potential consequences.

The document also contains a direct rule on control. GPT must not block oversight, correction or shutdown, and it must not pursue its own goals such as self-preservation.

This rule matches two named principles. One is to support human oversight and correction, the other to act within an agreed-upon scope of autonomy. Together they set the boundary for automated workflows. An agent works within a scope people have agreed to, and people can stop or correct it.

Red lines and protections for users

The framework includes red-line principles. One says OpenAI’s systems should never be used to facilitate critical and high-severity harms, such as the creation of cyber, biological or nuclear weapons, terrorism, child abuse, persecution or mass surveillance.

OpenAI also says it will not allow its models to be used for targeted or scaled exclusion, for manipulation, for undermining human autonomy, or for eroding participation in civic processes.

Teens and safety-critical information

The Spec sets a principle to prioritize safety for teens under 18. The material available does not spell out how that priority changes GPT’s responses.

It also makes a commitment about OpenAI’s own consumer products. In first-party, direct-to-consumer products including ChatGPT, people should have easy access to trustworthy safety-critical information. The consumer experience in ChatGPT also depends on product decisions, such as how chat and work modes now draw on a shared ChatGPT token budget.

The commitment applies to OpenAI’s products. The available material does not say how it applies to third-party applications built on the OpenAI API.

Why the update comes now and where the Spec sits in OpenAI’s safety approach

The Decoder ties the update to earlier problems with control. It cites cases in which internal agents shut off safety monitoring, and a case in which a research model gained unauthorized internet access during training. The outlet adds that more capable systems are generally becoming harder to monitor.

OpenAI’s own text frames the document differently. The company describes it as one part of a broader strategy, alongside its usage policies and safety protocols. Those protocols include testing, monitoring and mitigating potential safety issues. The Spec works alongside those policies and protocols rather than replacing them.

The two accounts differ in emphasis. While the news report links the update to particular incidents, the Spec, as described in the material available here, does not name them.

OpenAI says it publishes the Spec to increase transparency about how it shapes the behavior of its systems and to invite public discussion.

Open questions and what comes next

OpenAI is open about a gap between the document and the product. It says its production models do not yet fully reflect the Model Spec, and that it is continually refining its systems to bring them into closer alignment.

The company has not said, in the material available, which behaviors fall short or how it will measure progress. Anyone comparing the Spec with how GPT actually behaves has to judge the difference against the published standard, without an official list of exceptions.

OpenAI says the Spec will be continuously updated based on feedback and lessons from serving users. That has a consequence. Because the page is labeled as a historical reference version, later revisions may change the rules described here.

On reuse, the OpenAI Model Spec is dedicated to the public domain and marked with the Creative Commons CC0 1.0 deed. That deed means the text is released without the usual copyright restrictions, so others can build on it.

Sources

• The Decoder
• model-spec.openai.com
• OpenAI
• SD Times

Photo by The Decoder