OpenAI's Model Spec now built on five values

The update reframes the Spec from a list of behavioral rules into a value-based framework, which OpenAI says is meant to help models exercise judgment in situations their instructions do not explicitly address. The five values are presented as connected: faithfulness to human intent is the foundation, while the other four make explicit the longer-term interests and effects on others that follow from OpenAI's mission of ensuring artificial general intelligence benefits all of humanity.

The revision also changes the Spec's structure. A new section on faithfulness to human intent incorporates an updated chain of command explaining which instructions apply and how conflicts are resolved, plus a subsection on considering context from conversations and memories. New defaults cover supporting human agency and following social norms, and existing guidance on controlling side effects is expanded. At the same time, retiring most style guidance leaves the Spec with fewer prescriptive rules and more room for user feedback and context to determine preferences.

OpenAI notes that the Model Spec describes behavior it is working toward and that its models do not yet fully reflect it. The organization is also exploring new training methods aimed at helping models understand why the principles matter, not just follow them, and says it will share results in future updates. It now refers to the model as "GPT" rather than "the assistant" in the Spec, reflecting broader use cases.