OpenAI revised the Model Spec on Thursday—the public description of how it wants its models to behave. The new version is organized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
The reasoning is the more interesting part. Models increasingly act in situations their instructions do not cover, OpenAI writes, so interpreting instructions is not enough; what is needed is an understanding of the intent behind them. That is the same diagnosis Anthropic uses to justify its constitution, sorted differently.
Faithfulness is not obedience
The first value carries the other four. OpenAI writes the Spec and sets the boundaries, developers direct applications within them, and users direct the model’s work subject to those constraints. Faithfulness to intent is meant to be more than literal compliance: the model should understand what someone actually wants, and should expect that beliefs and values change—without quietly steering anyone in a direction.
The example in the text is an ordinary one. Comparing two job offers, the model might draw on priorities you shared earlier, such as predictable hours, while making that assumption visible and easy to correct. A new section, “consider context carefully,” governs how conversations, memories and other contextual evidence feed that understanding. Alongside it come defaults on human agency, on following relevant social norms, and on controlling and communicating side effects.
Style gets released, objectivity does not
The most concrete change is a deletion. Most of the Spec’s detailed style guidance is gone. Style should adapt to what people prefer, OpenAI says—in aggregate, within languages and cultures, and per person, with feedback and context as the evidence.
Two limits stay. Questions should still be approached objectively, and the model takes a particular perspective only when a user or developer asks for one. It may not infer that preference from what it believes the desired answer is. That is the anti-sycophancy line, and it sits next to the style release for a reason.
One small rename carries a signal: in the Spec the model is no longer “the assistant” but “GPT.” OpenAI puts that down to the breadth of deployment settings. An assistant sits next to you. A GPT runs somewhere.
A spec is a promise, not a description
The most honest sentence is at the end: the Model Spec describes the behavior OpenAI is working toward, and its models do not yet fully reflect it. In parallel the company is exploring training methods meant to teach these values rather than drill the rules, with results promised in future updates.
So the document describes an intention. That does not make it worthless—quite the opposite. There is now something to measure the next model against, in the vendor’s own words. It gets interesting the first time a model violates the Spec and OpenAI has to decide whether to fix the model or the Spec.