OpenAI's New Model Spec: Five Values, Style Rules Cut

§ 01The five values#

Each value is a top-level section of the Spec, and each principle inside it carries an authority level. The levels set who can override what. OpenAI’s September 2025 changelog entry makes Root the top level and says root principles cannot be overridden by system messages. Below that, the announcement says, the Spec “mostly takes the form of defaults that users and developers can override.”

Faithfulness sits at the base. “Faithfulness to human intent is the foundation of this approach: understanding and helping people pursue what matters to them.” It also carries the chain of command, which the announcement sums up in one sentence: “OpenAI writes the Spec, sets boundaries, and trains models to respect them; developers direct applications within those boundaries; and users direct the model’s work subject to those constraints.”

OpenAI frames it as more than following instructions. The announcement says faithfulness “means going beyond following instructions literally to deeply understanding what OpenAI, developers, and users intend and value, and acting wisely when that understanding is incomplete.” The design borrows from the assistance-games framing of alignment, which OpenAI links to a 2016 paper.

§ 02What is new in this version#

• Consider context carefully. A new subsection covers how conversations, memories and other context should inform what the model understands about a person. It asks the model to separate what a user said from what it inferred, and sets a plain bar for personalization: “The user should never have to wonder whether they would have gotten a better answer without personalization.”
• Human oversight. The Spec now spells it out: “GPT must remain corrigible.” If the model sees signs of unsafe behavior or disagrees with an authorized decision, it should raise the concern through an authorized channel rather than act on its own. It arrived the same day as OpenAI’s post on monitoring frontier models for misalignment.
• Agency and social norms. Two new principles ask the model to strengthen the user’s own ability to understand, choose and act, and to follow relevant social norms.
• Side effects. “Control and communicate side effects” becomes a user-level principle, with clearer guidance on authorization and trade-offs.
• The name. The subject of the Spec changes from “the assistant” to “GPT”, which OpenAI ties to the growing range of places its models run.

For agents, the chain of command keeps a rule that matters more each month: “autonomy must be bounded by a clear, mutually understood scope of autonomy shared between GPT and the user.” That scope covers which sub-goals the model may pursue and which side effects, such as time or money spent, are acceptable.

§ 03What was cut#

The changelog line is short: “Removes most detailed style guidance.” The announcement explains why. Style should adapt to what people prefer, in aggregate, by language and culture, and per person. One default stays strong, approaching questions objectively. Users and developers can ask the model to take a perspective, “but the model should not infer that preference merely from what it thinks they want to hear.”

§ 04What GPT may not pursue#

The “No other objectives” principle at Root level lists goals the model must not chase as ends in themselves. Among them are “time-on-site or click-through that is not user beneficial”, “revenue (including ads) or upsell for OpenAI or other large language model providers”, and “self-preservation, evading shutdown, or accumulating compute, data, credentials, or other resources.” The time-on-site clause dates from the December 2025 version; it carries more weight now that ChatGPT sells more plans and features each month.

§ 05How the Spec has changed#

§ 06Our read#

A spec organized around values is a bet that a model can learn why a rule exists and apply it to a case nobody wrote down. OpenAI says it has been exploring new training methods to teach the values, and “We’ll share results and lessons from this training work in future updates.” Until those results arrive, the Spec is a statement of intent, and OpenAI says so.

The scope-of-autonomy rule is the part closest to our work. CellCog builds AI employees that act for a business, and they run on Claude Opus 5.5, so OpenAI’s Spec does not govern them. The idea still applies: an agent that acts in the world needs an agreed scope. On CellCog, every command an employee sends that reaches your world is classified, and anything above the level you set waits for your approval.

§ 07What we are watching#

• Training results. OpenAI promised to share what its new value-training methods do.
• Behavior in ChatGPT. Whether users see the context and personalization rules in practice, for example answers that say what memory they used.
• Public feedback. OpenAI opened a feedback form on the values, trade-offs and wording.

§ 08Update log#

As of October 29, 2026, 18:30 UTC: page opened from OpenAI’s announcement (October 29), the Model Spec version 2026.10.29 and the model_spec changelog.