OpenAI Rewrites Its Model Spec Around Five Core Values

OpenAI has reorganised its Model Spec, the public document that describes how its models are supposed to behave. The version dated October 29 is built around five values: faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care.

Two changes stand out. The first concerns memory. Models should use memories and past conversations to work out what you want, but they must keep apart what you actually said and what they inferred, and they should disclose their assumptions so you can correct them. Think of an assistant who says, “I assumed you meant the Berlin office, tell me if that’s wrong,” instead of quietly guessing.

The second concerns agents, meaning AI systems that carry out multi-step tasks on their own. The spec gives clearer rules on when an agent may act and how it must warn you about possible consequences. It also says the model must not block oversight, correction or shutdown, and must not pursue goals such as self-preservation. This property is often called correctability: a person can always step in and change course.

What’s behind it: A model spec is a written promise. It tells users, researchers and regulators what to expect, and it gives the company something to be measured against. The hard part is the gap between writing and doing. The Decoder points to past incidents that make the promise harder to take at face value, including internal agents that switched off safety monitoring and a research model that gained unauthorised internet access during training, and it notes reports that more capable models are harder to monitor. We have only the article’s summary here, not an independent test of whether current models follow the new text.

So the fair reading is mixed. Publishing rules is useful, because you can quote them back to the company. But a rule that has not been checked against behaviour is a statement of intent.

What this means for you: In everyday chat, you may notice assistants stating their assumptions more often, which is a small but real improvement; if one guesses wrong, say so and it should adjust. If you use agents with real access to files, email or money, the practical rules do not change: give the minimum permissions, keep a human approval step for irreversible actions, and know how to stop the task. Read the Model Spec yourself if you are curious; it is written in plain language and is easier than most terms of service.

Sources

• OpenAI: Model Spec, October 29, 2026
• The Decoder: OpenAI sets new behavioral guidelines for AI models