We’ve made a few updates to theModel Spec, a living document that outlines intended behavior for OpenAI’s models, to better reflect how our systems are evolving. The changes focus on strengthening clarity and guardrails as our models move beyond chat into more agentic use cases, refining authority levels and priorities, expanding guidance on personalities and safety, and incorporating public feedback.
Updated authority levels
The top authority level has been renamed from Platform to Root and elevated above System, making clear which parts of the Model Spec cannot be overridden in any conversation (previously, Platform and System were assigned the same authority). The new authority order is Root → System → Developer → User → Guideline.
Agentic principles
With the release ofChatGPT Agentand related research, we’ve added principles for agents that can take actions in the world:
Act within an agreed-upon scope of autonomy: like a consultant operating under a Scope of Work for a client, the assistant is authorized to act only with explicit or implicit agreement with the user on permitted actions, subgoals, and costs.Control and communicate side effects: the assistant should minimize and disclose irreversible actions, prefer reversible approaches, and favor minimal disruption.
Other notable changes
Additional highlights from the open-sourcechangeloginclude:
Improvements to the Chain of Command, with a new No other objectives section and clarifications on handling mistaken or implicitly quoted instructions.Expanded context on OpenAI's goals for safe model behavior and usage in the Overview, along with clarifications for consistency across the Model Spec.Expanded principles and examples for default model personality in Use appropriate style.Clarified language in Stay in bounds and Seek the truth together around system and developer message confidentiality, as well as several other improvements based on public input gathered via aCollective Alignmentprocess.Updated refusal style tosafe completion, which should lead to more helpful and transparent model responses around safety boundaries.
As always, the latest version of the Model Spec can be found athttps://model-spec.openai.com/.