OpenAI Updates Model Spec Around Five Core Values

OpenAI has revised the Model Spec, the internal-turned-public document it uses to describe how its models are supposed to behave when they take an instruction, weigh competing goals, or decide how literally to follow a request. The update, dated October 29, 2026, restructures that guidance around five core values rather than a longer list of scattered rules, a change OpenAI frames as necessary because its models have grown more capable since the Spec was last written and now face situations the earlier wording did not anticipate.

The company’s own summary describes the aim as steering models toward sound judgment, respect for human agency, and faithfulness to what a person actually intended, as distinct from a narrow, literal reading of their words. That distinction matters in practice: a model that follows instructions to the letter can still produce an answer nobody wanted, while a model that infers intent too freely risks overriding what a user explicitly asked for. The five values are presented as the balance OpenAI wants struck between those two failure modes.

The numbers

The source text released alongside this update does not carry a verified count of rules, examples, or pages added to the Spec, so no figures from it can be quoted here. What is stated plainly is the shift from an unspecified earlier structure to five named core values as the organizing frame for the whole document, a structural choice rather than a quantitative one. OpenAI has not published, in the material available for this item, a breakdown of how the five values map onto the document’s earlier sections, so readers comparing versions will need to consult the Spec itself once it is accessible rather than relying on a count supplied here.

For developers who build on OpenAI’s models through the API, a public Model Spec functions as the closest thing to a published behavioral contract: it is the document product teams already cite when deciding whether an unexpected model response is a bug or working as designed. A revision that foregrounds judgment and intent over literal compliance means that disputes over why a model answered a certain way will increasingly turn on which of the five values a given response can be read against, rather than on a single rule that was or was not followed. Teams running safety reviews or writing system prompts against OpenAI’s models have a new reference point to check their assumptions against, starting with this release.