OpenAI gives the Model Spec five values and models more room for judgment

OpenAI published a new Model Spec on October 29. It now organizes intended GPT behavior around faithfulness to human intent, harm minimization, honesty and transparency, human agency and care.

Five values replace a collection of isolated rules

The new structure is meant to explain why individual rules exist and how to navigate conflicts between them. OpenAI added guidance on using context and memory, human oversight, corrigibility and control of side effects. The document also says GPT instead of assistant because the models now operate across a broader range of products and tasks.

Most detailed style guidance has been retired. Style should adapt to language, culture and user preferences. The chain of command remains broadly intact, with the highest Model Spec rules still taking precedence over developer and user instructions.

The agent should infer intent without taking the wheel

The practical tension sits between literal compliance and estimating what a person actually wants. GPT should use context, surface overlooked consequences and seek clarification for ambiguous or irreversible actions. At the same time, it should not treat its prediction of a user's future preference as a decision already made.

This matters most for agents with tools. The text gives more weight to budgets, side effects, the ability to stop an action and the user's ability to take over the work. For developers, the Model Spec increasingly resembles an operating policy rather than a chatbot tone guide.

Attractive values do not solve hard edge cases

OpenAI explicitly says current models do not yet fully reflect the Spec. Zvi Mowshowitz also argues that many new examples have an overdetermined correct answer. They show a desirable middle between two obviously bad extremes but offer little guidance for finding that middle in genuinely disputed cases.

Terms such as care, social norms and strengthening human agency can also point in different directions. Without evals that test conflicts among values, a model risks receiving an impressive map without a precise scale.

Evals must show how values change behavior

OpenAI will need to demonstrate the difference between the old and new versions on concrete scenarios. The most informative cases will pit an explicit instruction against a long-term preference, speed against reversibility and personal context against privacy.

The Model Spec is a public commitment and a useful object for criticism. Its weight will depend on measurable model behavior and whether users can understand why an agent acted, paused or refused to continue.

Lilith's verdict

OpenAI has drawn five cardinal directions for GPT. The compass becomes useful only when an agent can explain why it turned that way at a real junction.