OpenAI Rewrites Its Model Spec Around Five Values — and Drops Most Style Rules

OpenAI has published an updated version of its Model Spec, the public document that describes how the company wants its models to behave. The revision reorganizes the guidance around five stated values — faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care — adds new sections on context and oversight, and retires most of the document's detailed style guidance. OpenAI also says it is testing new training methods aimed at teaching models those values directly, but has not shared results.

What actually changed

The headline change is structural. OpenAI says the five values were already implicit across the Spec, and that grouping them explicitly helps explain why the document asks for the behavior it asks for, while exposing places where the commitments were underspecified.

Faithfulness to human intent is positioned as the foundation. OpenAI describes it as going beyond literal instruction-following to understanding what OpenAI, developers, and users intend and value, and acting sensibly when that understanding is incomplete. The company also says the model should recognize that people's beliefs and values change over time, and should help users reflect without quietly steering them. The example given is comparing job offers: the model might apply a previously stated preference such as predictable hours, but should surface that assumption so the user can correct it.

The other four values are described as making longer-term interests and third-party effects explicit. Harm minimization covers firm safety boundaries and foreseeable consequences of model actions. Honesty and transparency covers being truthful about what the model knows, does, and is uncertain about. Human agency covers strengthening people's ability to understand, choose, and act. Care covers attending to the circumstances of affected people, following relevant social and professional norms, and producing work people can rely on.

On the structural side, the Spec's new faithfulness section incorporates an updated chain of command explaining which instructions apply and how conflicts between them are resolved. A new subsection, "consider context carefully," covers how conversations, memories, and other contextual evidence should inform the model's read on user intent. OpenAI also added defaults on supporting human agency and following social norms, and expanded existing guidance on controlling and communicating side effects.

One naming change signals the direction of travel: the Spec now refers to the model as "GPT" rather than "the assistant." OpenAI attributes this to the growing breadth of use cases and deployment settings — a tacit acknowledgment that a document written for a chat assistant now has to govern agents taking actions in the world.

The style guidance retirement is the practical story

For anyone shipping on OpenAI's models, the most consequential line in the announcement may be the quietest one: most of the Spec's detailed style guidance is going away. OpenAI's reasoning is that style should adapt to what people prefer — in aggregate, within languages and cultures, and per individual — with user feedback and context serving as evidence of those preferences.

That is a defensible position, and it moves an area where a single global rulebook was always going to be a poor fit. It also moves style from a written, auditable specification to something determined by preference signals that are not public. Developers who relied on the Spec as a stable reference for tone and formatting defaults now have less to point at, and less ability to predict how defaults drift as aggregate feedback changes.

OpenAI appears aware of the obvious failure mode. The company says stronger defaults remain, including approaching questions objectively, and that while users and developers can explicitly ask the model to adopt a particular perspective, the model should not infer that preference from what it thinks the user wants to hear. That is a direct guard against sycophancy, which has been a recurring complaint about preference-trained assistants. Whether a written default holds up against optimization pressure from user feedback is an empirical question the announcement does not answer.

Assistance games, and what training is supposed to do

OpenAI says the faithfulness section borrows from the assistance-games framing in alignment research, in which the system treats human instructions as evidence about underlying preferences rather than as the preferences themselves. That framing is a reasonable fit for agents that encounter situations no instruction anticipated. It also raises the familiar objection: a system that infers what you "really" want has more room to be wrong in ways that are hard to detect. The Spec's emphasis on making assumptions visible and correctable is presumably meant to address that, though the announcement does not describe mechanisms for doing so in agentic settings where there may be no user watching a turn-by-turn conversation.

Alongside the Spec update, OpenAI says it has been exploring new training methods intended to help models learn these core values and apply the Spec's specific behavioral principles in unfamiliar situations — doing the right things, as the company puts it, for the right reasons. No results, methods, evaluations, or timelines were disclosed. OpenAI says it will share results and lessons in future updates.

The company is also explicit that the Spec is aspirational: it describes behavior OpenAI is working toward, and the models do not yet fully reflect it. That candor is useful, but it also means the document cannot currently be read as a description of deployed behavior. The announcement does not say which models, if any, have been trained against the updated Spec, nor does it describe how conformance is measured.

What's missing

The update is a governance document, not a product release, and should be read accordingly. Several things an enterprise buyer or regulator would want are not in the announcement: no compliance or evaluation data, no description of how the five values are weighted when they conflict beyond the existence of firm boundaries and overridable defaults, no detail on how developer-level overrides interact with the new agency and social-norms defaults, and no disclosure of how public feedback is collected, weighted, or incorporated. OpenAI says it wants feedback on the values, trade-offs, and specific language, and points readers to the full Spec.

The broader context is that model specifications are becoming a competitive and regulatory artifact. A published behavior standard gives customers something to audit against and gives regulators a document to cite. It also sets up a gap that will be measured: between what the Spec says models should do and what evaluations show they actually do. OpenAI has acknowledged that gap exists. The next useful disclosure is its size.

What to watch

• Whether OpenAI publishes evaluation results showing conformance to the updated Spec, and whether those evaluations are reproducible by outside parties.
• How the retirement of style guidance plays out in practice — particularly whether aggregate preference signals pull models toward agreeableness despite the stated objectivity default.
• Details on the new training methods OpenAI says it is exploring, including whether they generalize to situations the Spec does not explicitly cover.
• How the updated chain of command handles developer-versus-user conflicts in agentic deployments, where no human may be reviewing individual steps.
• Whether the rename from "the assistant" to "GPT" is followed by Spec guidance written specifically for long-running, tool-using agents rather than conversational turns.