OpenAI rewrites its Model Spec around five core values and retires most style rules
OpenAI updated its Model Spec on 29 October 2026, organizing it around five values, adding guidance on context and human agency, and retiring most style rules.
By Berkant A.Published
Tech AI LatestOpenAI published an updated Model Spec on 29 October 2026, the public document that describes how its models should behave. The new version is organized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It adds sections on using context and supporting human agency, retires most detailed style guidance, and calls the model GPT instead of the assistant. OpenAI says its models do not yet fully follow it.
Key facts
- Published: 29 October 2026 (OpenAI); full text at model-spec.openai.com
- Five organizing values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, care
- New guidance: consider context carefully, supporting human agency, following social norms; expanded rules on side effects
- Retired: most detailed style guidance; an objective point of view stays as a default
- The Spec now refers to the model as GPT rather than the assistant
- Not stated: which models or products are trained on this version, or results of the new training methods
What OpenAI changed
OpenAI published an updated Model Spec on 29 October 2026. The Model Spec is the company's public description of how it wants its models to behave, and the new dated version reorganizes the guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, facilitating human agency, and care. OpenAI says these values already ran through the document and that grouping them explains why the Spec asks for particular behavior.
Faithfulness to human intent is described as the foundation. OpenAI writes the Spec and sets boundaries, developers direct applications inside those boundaries, and users direct the model within both. The company says faithfulness means understanding what OpenAI, developers and users intend rather than following instructions literally, while not quietly steering people or overriding their choices. Its example: when comparing job offers, a model may use a priority the user shared earlier, such as predictable hours, but should make that assumption visible and easy to correct.
New and retired sections
The faithfulness section includes an updated chain of command, which decides which instructions apply and how conflicts are resolved, and a new subsection called consider context carefully. It covers how conversations, memories and other evidence should shape what the model infers about a person's intentions, with rules on carrying context forward, privacy and keeping context use visible and steerable. OpenAI also added defaults on supporting human agency and following social norms, and expanded guidance on controlling and communicating side effects of the model's actions.
At the same time OpenAI is retiring most of the detailed style guidance. It says style should adapt to what people prefer, in aggregate, by language and culture and per person, using feedback and context as evidence. Some stronger defaults stay, including taking an objective point of view: users and developers can ask for a particular perspective, but the model should not infer it just from what it thinks they want to hear. The Spec now refers to the model as GPT instead of the assistant, which OpenAI links to the wider range of products and deployments the models run in, from ChatGPT to the API.
Why it matters
The Model Spec is one of the few documents in which a frontier lab states in public what its models are supposed to do, and outside researchers and policymakers use it to judge model behavior against the company's own standard. Moving from detailed rules to values and judgment fits a broader shift in alignment work as models take longer, more autonomous actions where instructions run out. OpenAI says it is testing new training methods so that models learn these values and the reasons behind the rules, and will report results later. It also says plainly that the Spec describes a target: its models do not yet fully reflect it. More work on model behavior and evaluations is collected on the AI Research hub.
What to watch next
OpenAI invites public feedback on the values, trade-offs and wording. The things to watch are the promised results from the new value-training methods, whether future model cards say which Spec version a release was trained against, and how the retired style rules show up in everyday answers. Other labs have published comparable findings on agent behavior, such as Anthropic's report on unintended actions in live website tests.
What changed
| Before | The Model Spec gave rule-by-rule guidance, including detailed style rules, and referred to the model as the assistant |
|---|---|
| After | The Spec is organized around five values, adds guidance on context, oversight and human agency, drops most style rules and calls the model GPT |
| Who it affects | Users of ChatGPT and the OpenAI API, developers building on OpenAI models, and researchers who evaluate model behavior against the Spec |
Questions people ask
- What are the five values in OpenAI's updated Model Spec?
- Faithfulness to human intent, harm minimization, honesty and transparency, facilitating human agency, and care.
- Do OpenAI's models already follow the new Model Spec?
- Not fully. OpenAI says the Spec describes the behavior it is working toward and that its models do not yet fully reflect it.
- Why does the Model Spec now say GPT instead of the assistant?
- OpenAI says the change reflects the growing range of use cases and deployment settings for its models.
Sources
- An updated Model Spec for more capable and beneficial AI (OpenAI, 29 Oct 2026)
- Model Spec (2026/10/29) (OpenAI, 29 Oct 2026)
Reference pages
Spotted an error? See the corrections policy and tell the editor. Facts in this story come from the sources listed above; vendor claims are attributed.
Comments
Loading comments.
