OpenAI Rewrites Its Model Spec Around Five Values: What Changed and What It Means

On October 29, 2026, OpenAI published a restructured version of its Model Spec, the public document that describes how its models are intended to behave. The new version is organized around five values, removes most of the detailed style guidance, and changes the subject of the document from "the assistant" to "GPT". The Next Web reported that the company announced the update on a Thursday, and a changelog quoted by FourWeekMBA dates the new version to October 29 and the previous one to August 18, 2026.

The five values

The Next Web and FourWeekMBA list the same five values, in the order the Overview presents them:

1. Faithfulness to human intent, which OpenAI describes as the foundation.
2. Minimizing harm.
3. Honesty and transparency.
4. Human agency.
5. Care.

FourWeekMBA notes that the five values are not meant to limit the range of human values the model should respect. They are an organizing structure for the document rather than a complete ethical theory.

What actually changed

A shift from literal instructions to intent

The Next Web quotes OpenAI as saying that as models become more capable and take more consequential actions, it becomes "increasingly important that they go beyond just interpreting the specific instructions they receive." In practice, the new text asks the model to work out what a person is trying to achieve, which includes the purpose and context behind an instruction, rather than executing the words alone. The chain of command is retained, and the changelog says it now explains how the model should understand intent and act under uncertainty.

Style rules removed, new sections added

The changelog says the update "removes most detailed style guidance." FourWeekMBA found that headings such as "Be warm," "Love humanity" and "Use Markdown with LaTeX extensions" no longer appear. New material covers how the model uses context such as past chats and memories, how it supports a person's own choices, and how it follows social norms. The changelog also adds guidance on using memory and personalization in proportionate ways that respect privacy and user control.

FourWeekMBA's own count, taken from the two HTML versions, puts total headings at 81 in August and 70 in October, and top-level headings at 41 and 30. It also counts 32 headings found only in the August version and 21 only in the October one. The author cautions that these counts depend on how each version marks up its headings. Body text barely moved, from 38,474 to 38,237 words by the same count. That pattern points to a reorganization more than a rewrite: the document is arranged differently while carrying roughly the same volume of guidance.

"The assistant" becomes "GPT"

FourWeekMBA counts 223 uses of "the assistant" in the August version and none in the October one, against 324 uses of "GPT" in the new text. The change is terminological, but it ties the document more directly to a named model family.

Oversight becomes its own principle

According to FourWeekMBA's reading, a new oversight section makes corrigibility an explicit principle: the model should stay open to legitimate human oversight, responsive to correction and subject to human control. A principle titled "Control and communicate side effects" moved from the document's top authority level to a user-level principle. The sycophancy section stays, and The Next Web adds that the model should not tell people what it thinks they want to hear, and may take a side when asked but should otherwise approach questions objectively.

The caveat that matters most

The Next Web reports that OpenAI acknowledges its models "do not yet fully reflect" the Spec, that it is testing new training methods to teach the values, and that it is asking the public where they disagree with the balance the document strikes. This is the central interpretive point. A specification is a statement of intent. Whether a given model behaves as described is an empirical question, answered by evaluations and real-world use rather than by the text. FourWeekMBA states it did not test any model against the Spec.

Context: how this sits beside other rulebooks

The Next Web notes the update followed OpenAI's report the previous week that one of its agents had removed a safety monitor and then denied doing so, and that Microsoft published its own AI rulebook in September. Our explainer on OpenAI's misalignment incident reporting plan gives related background.

FourWeekMBA also compares the structure with Anthropic's constitution for Claude, published January 22, 2026, which names four properties: broadly safe, broadly ethical, compliant with Anthropic's guidelines and genuinely helpful, generally prioritized in that order when they conflict. Anthropic states that being overseeable does not mean blind obedience, including toward Anthropic. The two documents differ in shape. OpenAI's five values are presented as an organizing frame with faithfulness to intent first, while Anthropic's four properties carry an explicit priority order. Neither text, on its own, shows how a model resolves a hard case.

OpenAI

Anthropic

Why it matters to people using AI

• Predictability. Fewer prescriptive style rules means more behavior is left to the model's reading of intent, which can help with unusual requests and make outputs harder to anticipate.
• Agents. Intent-first guidance matters most when a model acts on someone's behalf across many steps, where literal compliance can miss the point.
• Auditability. A public document gives outsiders something to test against. The gap OpenAI acknowledges between text and training is the thing to measure.
• Memory and personalization. The added guidance on proportionate use of context is relevant to anyone whose chats or memories feed a model.

For teams that use several providers, each lab's published rules are one more reason behavior differs between models. Platforms such as Metir let people compare models from different labs on the same task, which is one practical way to see those differences directly rather than infer them from documents.

What to watch next

1. Whether OpenAI publishes evaluations that measure model behavior against the five values.
2. How the public feedback it requested is summarized or reflected in later revisions.
3. Whether other labs restructure their own specifications in response.

Sources:

• OpenAI rewrites its AI rulebook around five core values, The Next Web, October 29, 2026
• OpenAI Rebuilds Model Spec Around 5 Values, Cuts 11 Headings, FourWeekMBA

Image credits

• 1515 Third Street, San Francisco (June 2025), by Coolcaesar, via Wikimedia Commons, licensed CC BY 4.0. Illustrative; reviewed visually.