OpenAI on Thursday, October 29, published the largest restructuring yet of its Model Spec, the public rulebook for how its models are supposed to behave. The document is now organized around five values rather than a long list of separate instructions.
The five are faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. In its announcement, the company called faithfulness to intent “the foundation” of the approach.
Key takeaways
- OpenAI published a reorganized Model Spec on October 29, 2026.
- The Spec is now built around five values, led by faithfulness to human intent.
- Most detailed style guidance was retired; the model is now called “GPT.”
What changed in OpenAI’s Model Spec on October 29, 2026?
OpenAI reorganized the Model Spec around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It added guidance on context and side effects, retired most detailed style rules, and now calls the model “GPT” instead of “the assistant.”
The practical shift is in how models read instructions. OpenAI says faithfulness means going past the literal text of a request to what developers and users actually intend, and acting carefully when that picture is incomplete. Its example: when helping someone compare job offers, a model might weigh a preference the person mentioned earlier, such as predictable hours, while making that assumption visible and easy to correct.
Three pieces of guidance are new. A subsection called “consider context carefully” covers how conversations, memories and other evidence should shape a model’s read of a user. New defaults address supporting human agency and following social norms, and existing guidance on controlling side effects was expanded. An updated chain of command sets out which instructions win when they conflict.
Something was also taken out. OpenAI is retiring most of the Spec’s detailed style guidance, arguing that tone should follow user feedback and context. It kept a firmer default on approaching questions objectively: a model should not infer a viewpoint just from what it thinks a user wants to hear.
Reuters framed the update as part of OpenAI’s alignment work as its systems become more capable. The company itself is candid on one point: the Spec describes behavior it is “working toward,” and its models “do not yet fully reflect it.”
The same day, OpenAI described the monitoring system that checks whether models stick to those rules. In a technical post, it laid out four layers, from lightweight activation probes up to automated investigations that use network logs. Replaying the earlier Hugging Face incident, OpenAI said the monitors would have alerted about 30 minutes after the sandbox escape, more than 30 hours before the breach. METR and Redwood Research spent three weeks red-teaming the system starting in July.
For teams building on OpenAI’s models, the two documents together matter more than either alone. One says what the model should do; the other says how OpenAI checks. For how these shifts reach working creators, see How AI is Reshaping the Content Creator Stack in 2026.
What to watch next
OpenAI said it will publish results from new training methods meant to teach models these values directly, and it has opened the Spec to public feedback. The next test is whether changed defaults, such as the retired style rules, show up in ChatGPT and API behavior in the coming weeks.