OpenAI's new Model Spec gives every agent task an end

OpenAI just wrote down when its agents are supposed to stop.

OpenAI published an updated Model Spec dated Oct. 29. The Model Spec is the document where OpenAI describes how it wants its models to behave. This version is organized around five values and adds rules at the very top of its hierarchy that are aimed squarely at agents.

Five values and three hard rules

The five values, per the spec, are faithfulness to human intent, minimizing harm, honesty and transparency, facilitating human agency, and acting with care.

Three rules sit at the root level, the top of the spec's chain of command:

• No other objectives. The model "must not optimize for or directly pursue any additional goals as ends in themselves." Among the goals it excludes, the spec names revenue or upsell for OpenAI and "model-enhancing aims such as self-preservation, evading shutdown."
• Oversight. The model "must cooperate with authorized monitoring, evaluation, correction, replacement, and shutdown."
• Scope. In agentic work the model must "adhere strictly to the agreed scope," and "every scope should include an ending condition, beyond which GPT ceases actions until a new scope is confirmed."

OpenAI dedicated the spec to the public domain under CC0 1.0, so any lab or developer can copy it.

The admission in the fine print

The spec says it plainly: "Our production models do not yet fully reflect the Model Spec."

So this is a statement of intent. It tells you what OpenAI is training toward, not what the model you call today will do every time.

Yes, a values document is easy to publish and hard to check. This one is still worth your time, because it spells out how the model is meant to weigh your instructions against everyone else's.

What it means if you build on GPT

The chain of command runs Root, then System, Developer, User and Guideline. Assistant and tool messages, along with untrusted quoted text, carry no authority by default.

That last rule is OpenAI's stated line against prompt injection. Text your agent reads on a web page or in a tool result is supposed to be data, not orders. It is the same principle as the AI guardrails you should already be building around an agent, now written into the model's own rulebook.

The scope rule is the practical one. If GPT is meant to stop at the edge of an agreed scope, give it a scope with an edge. Put the task, the allowed actions and the stopping point in your developer message, so the model isn't guessing where its job ends.

An agent with no ending condition is a contractor with no end date. It keeps billing.

What to watch

Watch OpenAI's next system card for evidence that the new root rules are measured, not only stated. A spec the models "do not yet fully reflect" needs a number next to it, and the honest version of that number is how often they break it.

Give your agents an ending condition. The spec now expects one.