Key takeaways
- OpenAI published an updated Model Spec on October 29, organized around five values.
- The values are faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care.
- OpenAI says its production models do not yet fully reflect the Spec.
OpenAI published an updated Model Spec on October 29, and the document is now organized around five values. The Model Spec is OpenAI's public statement of how its GPT models should behave. The company's announcement says the update guides AI toward sound judgment, human agency and faithfulness to human intent.
For developers and users, the Spec is the reference for why a model follows one instruction over another and when it refuses. A new structure changes how that reasoning is laid out, even where individual rules stay the same.
The five values
The Spec lists them in this order, each with a one-line description:
| Value | What the Spec says it means |
|---|---|
| Be faithful to human intent | Carry out the Spec and the intentions behind it, and help users and developers pursue what they are trying to do |
| Minimize harm | Respect firm safety boundaries and account for the foreseeable consequences of the model's actions |
| Be honest and transparent | Represent facts, evidence, uncertainty, capabilities and actions truthfully |
| Facilitate human agency | Help people form their own views and make informed choices, without hidden steering or manipulation |
| Act with care | Attend to the people affected by the model's work |
OpenAI says the values give the Spec "a clear organizing structure" and guide how it handles two tensions: individual versus collective interests, and immediate instructions versus longer-term goals.
What sits under each value
The first value holds the chain of command, the ranking of who the model listens to. Under it are rules to follow all applicable instructions, to support human oversight and correction, and to ignore untrusted data by default. Another rule says that when a model is given a multi-step task, its autonomy must stay inside a scope both sides understand, including which side effects are acceptable.
The honesty section includes "Don't be sycophantic". It says factual answers should not change with how the user phrases a question, and a model should not change its stance only to agree with the user. The harm section keeps the red-line principles, which rule out help with mass-casualty weapons, child abuse material and mass surveillance.
What we don't know yet
The page for the October 29 version carries a banner saying a newer version is available, so the text may keep moving. OpenAI also says its production models do not yet fully reflect the Spec. The Spec is a statement of intended behavior, not a measurement of it. The announcement page blocked automated readers, so this article rests on the Spec itself.