OpenAI has published a version of its Model Spec, dated 29 October 2026. The document describes how its GPT models are meant to behave. OpenAI says it is built around five values. It sets a chain of command, so instructions with higher authority win over lower ones. OpenAI says most of its rules are defaults that users can change. The firm lists a smaller set of limits, such as avoiding catastrophic risks and not exposing private personal information.
OpenAI says its production models do not yet fully reflect the spec. It says it is still refining its systems to bring them closer. The page is marked as a historical version that may not match current policy. This overview does not say what changed from earlier versions. OpenAI also says the public text may leave out some details, though it calls the text consistent with its intended model behavior.
For teams that build on OpenAI's API, the spec shows who can change what. Developers can override some defaults that users cannot. But OpenAI itself says live models fall short of the document. Builders should read it as a statement of intent, not a guarantee of behavior.