OpenAI Model Spec update: five values, less style guidance

OpenAI published an updated Model Spec on October 29. The Model Spec is OpenAI's public description of how its models should behave, and this update reorganizes it around five values, adds guidance on context and oversight, and retires most of the detailed style guidance. If your support bot, quoting tool or internal agent runs on GPT, this is the rulebook your system prompt sits inside. Two changes matter for operators: the model is told to read intent, not just instructions, and it no longer gets a fixed house style from OpenAI.

What actually happened

Per OpenAI's announcement:

• Five values. The Spec is now organized around faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. OpenAI calls faithfulness to human intent "the foundation."
• A layered chain of command. OpenAI sets the boundaries, developers direct applications inside them, and users direct the work inside the developer's constraints. The update includes a revised section on which instructions apply and how conflicts get resolved.
• Context counts. A new subsection tells the model to use conversations, memories and other context to understand what people want.
• Style guidance is mostly gone. OpenAI says style should adapt to what people prefer, based on feedback and context. Some stronger defaults stay, such as approaching questions objectively.
• New defaults on supporting human agency and following social norms, plus expanded guidance on controlling and communicating side effects.
• "GPT," not "the assistant." The Spec now names the model GPT.

OpenAI also says plainly that the Spec "describes the behavior we are working toward" and that its models do not fully reflect it yet.

Why the Model Spec update matters for your business

Your system prompt is now a statement of intent. A model trained to look past the literal instruction will fill gaps with what it thinks you meant. That is good when your prompt explains the goal. It is a risk when your prompt is a list of rules with no reason attached. Say why: "Never quote a delivery date — our carrier changes ETAs daily" gives the model something to reason from when a customer asks a new question.

Put your voice in writing. With OpenAI's own style rules retired, tone will drift toward whatever the model infers from context. If your brand has a voice, spell it out in the developer instructions. Do not assume last quarter's defaults.

Side effects need a rule from you. The Spec expands guidance on communicating side effects. For an agent that edits orders or sends emails, decide which actions need a confirmation step and enforce that in code, not only in the prompt.

Re-run your evals. The Spec is a target, and training follows it over time. Keep a fixed set of real customer questions and score each model update against it before it reaches production.

Sources: OpenAI.