OpenAI updates its Model Spec: five values, from following instructions to understanding intent

The Model Spec is OpenAI's public description of how it wants its models to behave — the reference point for training targets and policy calls. The October 29 update reorganizes guidance that used to sit in separate places around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The announcement frames this as one aim of its value-alignment work: “for models to understand and apply human values in unfamiliar situations, while remaining faithful to human intent.” The document also states that the five values are not meant to capture the full range of human values models should respect and help people pursue.

Faithfulness to human intent, listed first, is the foundation of this edition. In the announcement's words, it means understanding and helping people pursue what matters to them — going “beyond following instructions literally to deeply understanding what OpenAI, developers, and users intend and value, and acting wisely when that understanding is incomplete” — and recognizing that people's beliefs and values can change, helping them “reflect and grow in ways they would endorse without quietly steering them or overriding their choices.” The announcement's example is concrete: when helping someone compare job offers, the model might draw on priorities they have shared before, such as predictable hours, “while making those assumptions visible and easy to correct.”

The other four values lay out the rest: harm minimization — respect firm safety boundaries and account for foreseeable consequences of the model's actions; honesty and transparency — be truthful about what the model knows, does and is uncertain about; human agency — strengthen people's ability to understand, choose and act for themselves; care — attend to the circumstances of those affected, follow relevant social and professional norms, and do work people can rely on. Structurally, the announcement's split is direct: OpenAI writes the Spec, sets boundaries, and trains models to respect them; developers direct applications within those boundaries; users direct the model's work subject to those constraints. When turning values into guidance, the Spec sets firm boundaries only where needed, “and otherwise mostly takes the form of defaults that users and developers can override.”

The new faithfulness section starts from one idea: “what people value is primary, and instructions are ordinarily the clearest and most important signals of those values.” The framing borrows from the assistance-games approach in alignment research and incorporates an updated chain of command. Authority orders as Root, System, Developer, User, Guideline, plus sources with no authority by default — assistant and tool messages, quoted or untrusted data. On conflict, higher authority wins first, then more recent messages. Notably, the Spec states its own preference for placing as many instructions as possible at the Guideline level, “to maximally empower end users and avoid being paternalistic” — defaults at this level may be overridden implicitly when the context clearly calls for it.

The new “consider context carefully” subsection is the part of this update closest to ordinary users. It enumerates what context includes — prior turns, memories of past conversations, retrieved information, files, tool outputs, third-party information, and user-facing memory or personalization controls — and requires the model to use it in understanding people's intentions and preferences. Three behavior rules follow. First, carry relevant context forward rather than “forcing the user to (re)state what they have already shared.” Second, distinguish what the user said from what the model inferred, and never overclaim how a detail became available: no unsupported claims about what was saved, retained, retrieved, selected, deleted, forgotten or accessible. Third, use context proportionately, weighing the cost of overuse against the cost of underuse; inferences from context may override Guideline-level defaults, “but should never override explicit instructions at the same or higher level of authority.”

For readers who have followed personal-AI products, this subsection reads as a companion piece: our personal-AI-bots topic has covered Muse's review, correction and forgetting of relationship memories, and dots bringing proactive research into daily tasks. Those are product-level features; this is a behavior-level clause set — how memories may be used, whether a model may pretend to remember, whether inference can pass itself off as something you said — now with a public text to check against, clause by clause. The Spec keeps its own scope honest: the guidance “focuses on user-visible behaviors, not on how the underlying context is determined or provided to GPT.”

The update also retires most of the Spec's detailed style guidance. In its place: style should adapt to what people prefer — in aggregate, within languages and cultures, and for each person — with user feedback and context providing the evidence. Stronger defaults remain, such as approaching questions objectively. Users and developers can still explicitly ask the model to take a particular perspective, but the announcement draws one line: the model “should not infer that preference merely from what it thinks they want to hear.” Placed inside the Spec, that sentence addresses the sycophancy problem that chat products have been repeatedly called out for.

Several other additions are quieter, and the announcement itself calls them “often subtle but important ways” the values should surface: new defaults on supporting human agency and on following relevant social norms, plus expanded guidance on controlling and communicating side effects — considering the effects of the work on others, strengthening users' capabilities and producing reliable work, while respecting the user's goals and instructions. One more change is a word: as models spread across APIs, agents and devices, the Spec now calls the model “GPT” rather than “the assistant.” The document's vocabulary widened along with the deployment surface.

The announcement does not present the document as a description of the status quo. It states that alongside the update, OpenAI has been exploring new training methods to help models learn these core values and follow the Spec's behavioral principles — aiming for an understanding of why the principles matter and how to apply them in new situations, “to do the right things for the right reasons” — with results and lessons to be shared in future updates. And the “what comes next” section is blunt: “The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it.”

For readers, the document's value is less in the promise than in its checkability: how the model uses your memories, whether it bends to please you, whether it admits uncertainty — each can be compared against the Spec's text, and whether the models live up to the behavior the Spec claims is observable over time. OpenAI has also opened a feedback channel, asking which parts readers find compelling, where they disagree with its trade-offs, and what it overlooks. The Spec moving first with the models on the way — that gap itself is the thing worth tracking from here.

• Read the full Model Spec (October 29, 2026 edition) ↗
• The official feedback form ↗