OpenAI replaces the Model Spec’s style checklist with five values
On October 29, 2026, OpenAI reorganized its Model Spec around five values — faithfulness to human intent, harm minimization, honesty, agency, and care — and made them the foundation of an explicit chain of command. For teams that relied on the models’ default tone, the consequence is direct: style now has to be specified, because most of the detailed style rules have been retired.
October 29, 2026. OpenAI publishes a major update to its Model Spec, the public document describing how its models should behave. Five values. The old organization by detailed rules gives way to a framework built around five principles. “GPT.” The model is no longer called “the assistant” in the document. The point that matters for an operator: most of the detailed style rules are retired, which shifts responsibility for tone and format onto whoever writes the prompt.
Five values, one foundation
The reorganization is built around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. OpenAI presents them as commitments already reflected across the Spec, now made explicit to explain why the document asks models to behave as it does.
Faithfulness to human intent is positioned as the foundation. It goes beyond executing instructions literally: the model should understand what OpenAI, developers, and users actually intend, and act with judgment when that understanding is incomplete. The example given is concrete — when helping someone compare job offers, the model might draw on previously shared priorities such as predictable hours, while making that assumption visible and easy to correct.
The other four values spell out the counterweights. Harm minimization means respecting firm safety boundaries and accounting for foreseeable consequences. Honesty and transparency mean being truthful about what the model knows, does, and is unsure of, so people can evaluate its reasoning. Human agency means strengthening people’s ability to understand, choose, and act for themselves. Care covers attending to the circumstances of those affected and following relevant social and professional norms.
A chain of command that makes authority explicit
The heart of the change is a formalized chain of command. OpenAI writes the Spec, sets boundaries, and trains models to respect them; developers direct applications within those boundaries; users direct the model’s work under those constraints. It is an authority hierarchy that previously existed only implicitly, now written down — and it settles how conflicting instructions are resolved.
The structure borrows from the assistance games framing of alignment, cited in the document. A new “consider context carefully” section explains how conversations, memory, and other contextual evidence should inform the model’s understanding of user intent — a sensitive point as models retain ever more history.
OpenAI also says it is exploring new training methods to teach these values “more directly,” with an explicit goal: to do the right things for the right reasons. The Spec, the company stresses, describes a target behavior — current models do not yet fully reflect it.
What disappears: the style rules
The most concrete consequence for a user is not philosophical, it is practical. The Spec retires most of its detailed style rules. The reasoning is stated plainly: style should adapt to what people prefer, using their feedback and context as evidence. The document keeps a few stronger defaults, such as “approach questions objectively” — a user can ask for a particular perspective, but the model should not infer that preference from what it thinks the user wants to hear.
A vocabulary change accompanies it all: the model is now called “GPT” rather than “the assistant,” reflecting the breadth of deployment settings, from consumer chat to autonomous agents.
For teams building on these models, the effect is direct. An application that leaned on the model’s default punctuation, format, or tone loses that safety net: those behaviors are no longer guaranteed by the Spec and must be spelled out in the system prompt. It is a transfer of responsibility — from the lab to the developer — consistent with the logic of the chain of command.
Two tensions the Spec must arbitrate
The document does more than list values: it spells out the two tensions they exist to resolve. The first pits individual interest against collective interest — helping a user pursue their goals without ignoring the effects of that help on others. The second pits immediate instructions against longer-term goals: follow what the user asks for right now, or protect what they want durably, including against themselves. Neither fracture resolves with a rule; both resolve with judgment — and that is precisely what the five values are meant to guide.
The history helps read the move. This is the fifth revision of the Model Spec in fourteen months, after the September 2025 version (“the root above the system”), October 2025 (tool-output delegation), December 2025 (under-18 mode), and August 2026 (capabilities and limits). The direction is clear: from a checklist of behaviors toward a value framework meant to hold up as models gain autonomy and take more consequential actions in the world.
For engineering teams, the consequence plays out in the system prompt. Retiring the detailed style rules does not mean the model has no tone: it means you have to define it. An application that specifies neither format, nor perspective, nor level of caution now inherits a less predictable behavior than before — not through OpenAI’s negligence, but through a deliberate choice to delegate style to the user. The Spec becomes a governance document, not a behavior contract, and that is a shift many teams have not yet priced in.
Underlying the rewrite is a broader research program OpenAI calls value alignment: teaching models to understand and apply human values in unfamiliar situations, rather than handing them an encyclopedia of rules. It is an answer to the fundamental problem of behavioral specs — a rule covers the cases you anticipated, a value covers the ones you did not. The trade-off, which OpenAI accepts, is that a value framework is harder to audit: you do not verify a value the way you verify a rule, and the document’s own caveat — that models do not yet fully reflect it — makes that gap explicit.
For evaluators, two of the five values deserve particular attention. Honesty now extends to what the model is uncertain about, not just what it knows — a concrete, testable behavior. And care imports social and professional norms into the defaults, which means the model’s behavior can now legitimately vary by context in ways a single test harness may not capture. Both are signals that the Spec has stopped being a style guide and started being a liability document. Read that caveat alongside the retired style rules and the picture is clear: the defaults you once inherited are now yours to write, and a team that treats the Spec as a finished behavior contract will be the first to be surprised by a model that has not yet caught up to the document.
Verdict
If you build products on OpenAI’s models, stop counting on implicit style defaults: lock tone, format, and perspective into your system instructions, because the Spec no longer guarantees them. If you run AI governance or compliance, read the “chain of command” and “context” sections — that is where the trade-offs between immediate intent and long-term interest, and between the individual and the collective, now play out. And if you are evaluating a model’s behavior, keep OpenAI’s own caveat in mind: the Spec describes a target, not the current state of the models — measure actual behavior, not the document that describes it.