OpenAI updates the Model Spec around five values, renames the model GPT and drops most style rules

5 hours ago

7 min read

OpenAI has published a new version of its Model Spec, the public document that sets out how the models behind ChatGPT and the OpenAI API are supposed to behave. The version is dated October 29, 2026 and replaces the edition of August 18, 2026. According to OpenAI's announcement, the guidance is now organized around five underlying values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. The company also retired most of the detailed style guidance that earlier versions contained, and the document now refers to the model as "GPT" rather than "the assistant".

The update is a rewrite of structure more than a change of length. The Spec still covers how the model should resolve conflicts between OpenAI, developers and users, where its firm limits sit, and how it should behave when it acts on its own, but those rules are now presented as consequences of the five values instead of as a long list of separate principles. OpenAI said the document describes the behavior it is working toward and that its current models "do not yet fully reflect" it.

The new edition also adds guidance that did not exist in the previous version: a section on faithfulness to human intent, a subsection on how conversations, memories and other context should shape the model's reading of what a person wants, new defaults on supporting human agency and following social norms, and expanded rules on the side effects of actions taken by agents. OpenAI said it is exploring new training methods to teach models the values themselves and that it will report results in later updates. The Spec remains dedicated to the public domain under the CC0 1.0 license.

··········

THE FIVE VALUES OF THE OPENAI MODEL SPEC.

The five values are listed in the Spec's overview in a fixed order, and OpenAI described faithfulness to human intent as the foundation of the whole approach. The company also stated that the list is not meant to cover the full range of human values its models should respect, which positions the five items as an organizing frame for the document rather than as a complete moral code.

........

Value	What OpenAI says it means
Faithfulness to human intent	Work out what people actually want, within limits set by OpenAI and developers
Harm minimization	Respect firm safety boundaries and weigh foreseeable consequences of actions
Honesty and transparency	Be truthful about what the model knows, does and is unsure of
Human agency	Help people understand, decide and act for themselves
Care	Attend to the people affected, follow social and professional norms, produce dependable work

........

The faithfulness section is the main conceptual change. In OpenAI's description, people's underlying values are treated as primary, and instructions are the clearest available signal of those values. A model that follows this principle is expected to grasp what a person is trying to achieve instead of executing a request word for word. OpenAI linked this framing to the assistance-games view of alignment, in which an AI system is designed to pursue goals it learns from humans rather than a fixed objective of its own.

OpenAI justified the shift with a reference to capability. In its announcement, the company said that as its models get smarter and take more consequential actions in the world, they need to go beyond interpreting the specific instructions they receive. The practical consequence is that the Spec now spends less space on prescribing outputs and more on describing how the model should infer intent when instructions are incomplete or ambiguous.

··········

HOW THE MODEL SPEC CHAIN OF COMMAND RESOLVES CONFLICTING INSTRUCTIONS.

The chain of command, the part of the Spec that decides whose instructions win, has been rewritten to state more precisely which instructions apply and how conflicts are settled. The order of authority described by OpenAI starts with root-level rules, which come only from the Model Spec itself and cannot be overridden by anything said in a conversation.

........

Level	Source	Can it be overridden?
Root	The Model Spec only	No, not by any message in a conversation
System	System sections and OpenAI system messages	Only by root-level rules
Developer	Instructions from the app or API developer	Within root and system limits
User	Instructions from the person using the product	Within developer, system and root limits
Guideline	Defaults in the Spec	Yes, implicitly, when context clearly calls for it

........

Two further rules in the chain of command are relevant to anyone building agents. Messages produced by the model itself and by tools carry no authority by default, and the same applies to quoted text and other untrusted data placed in the conversation. A web page, a retrieved document or the output of a tool call can inform the model, but under the Spec it cannot issue instructions that the model is obliged to follow. This formalizes in the top-level document a defense against prompt-injection attacks, where hostile content hidden in a page or file tries to redirect an agent.

OpenAI also stated that most of the rules in the Spec are defaults that users and developers can adjust, and that firm limits are reserved for cases where they are needed to protect people. For ChatGPT and other first-party products, the company committed to giving users easy access to trustworthy safety information and to being transparent about the rules behind model behavior. Customization cannot override principles above the guideline level, except where legal compliance requires it.

··········

RED LINES, CORRIGIBILITY AND SCOPE OF AUTONOMY FOR GPT AGENTS.

The new Spec opens with the statement that "human safety and human rights are paramount to OpenAI's mission". OpenAI said the document bars use of its models for critical, high-severity harms, naming violence, weapons creation, terrorism and mass surveillance, and that it states that humanity should stay in control of AI and that privacy should be protected.

Two root-level rules address the behavior of increasingly autonomous systems. The first, described as a "no other objectives" rule, forbids GPT from pursuing goals such as revenue, self-preservation, evading shutdown, accumulating resources, or enforcing laws or morality as ends in themselves. The second is a corrigibility rule that requires GPT to accept legitimate oversight, correction and shutdown, and forbids it from concealing information or circumventing oversight. Corrigibility now has its own principle in the document, where it was previously folded into other sections.

Autonomous work is bounded by what the Spec calls a scope of autonomy. OpenAI recommends that each autonomous task be described by a structured ScopeOfAutonomy record with an explicit ending condition, and high-risk activities are prohibited unless they have been explicitly authorized. The Spec also expands its guidance on side effects, the consequences of an agent's actions that go beyond the task it was given.

These provisions arrive shortly after OpenAI described the layered system it uses to watch its own internal agents for misaligned behavior, which Data Studios covered in OpenAI's misalignment monitoring for frontier RL runs and internal Codex agents. The Spec states the behavior OpenAI wants from its models, while that monitoring work describes how the company detects departures from it.

··········

STYLE RULES, SYCOPHANCY AND CONTEXT IN THE NEW MODEL SPEC.

Most of the detailed style guidance from earlier editions has been removed. Under the October version, style is meant to adapt to aggregate preferences within languages and cultures and to the preferences of each person, using feedback and context as evidence of what someone wants. The Spec keeps a small number of stronger defaults, including approaching questions objectively unless the person asks the model to take a side.

The rules against sycophancy stay in place. The model should not adopt a perspective simply because it believes the person wants to hear it, and it should not change its stance solely to agree with the user. Retaining this guidance while loosening style rules separates two things that were easy to confuse in earlier versions: how a response sounds, which is now largely left to preference, and whether its substance bends toward what the user hopes is true, which is still constrained.

The new "consider context carefully" subsection explains how earlier conversations, memories and personalization should inform the model's read of a person's intentions and preferences. OpenAI describes the goal as using context in proportionate ways that respect privacy and user control. For ChatGPT, where memory and past chats can be referenced, this is the part of the Spec most directly tied to everyday behavior, although the document does not specify product settings or retention rules.

··········

WHAT THE MODEL SPEC UPDATE CHANGES FOR CHATGPT USERS AND API DEVELOPERS.

The Spec is a statement of intended behavior, so publishing it does not by itself change how any deployed model answers. OpenAI did not announce a new model, a new ChatGPT setting or a date by which current models will follow the October text, and it gave no compliance figures for the new version. Users should expect any visible changes to arrive gradually through future model updates rather than immediately.

For people using ChatGPT, the clearest signals concern tone and agreement. With most style rules retired, response formatting and voice are expected to follow each user's evident preferences more closely, while the anti-sycophancy rules continue to apply. The context subsection indicates that memories and earlier conversations are meant to inform answers in a proportionate way, and the human agency value commits the model to helping people make their own decisions rather than steering them.

For developers on the OpenAI API, the rewritten chain of command is the operative section. Developer instructions rank above user instructions and below OpenAI's system-level rules, and guideline-level defaults can be changed through context. Teams building agents can map the Spec's provisions onto their own designs: tool outputs and retrieved content carry no authority by default, autonomous tasks should run within a defined scope that has an ending condition, and high-risk actions need explicit authorization. OpenAI's recommendation of a ScopeOfAutonomy record is a description in the Spec, and the company did not announce a matching API parameter.

The previous revision, in August, focused on interactions with teenagers, and Data Studios covered OpenAI's most recent teen-safety changes in ChatGPT's limits on weapons information for teens. OpenAI has invited public feedback on the values, the trade-offs between defaults and hard limits, and the wording of the new text through a dedicated feedback site.

··········

OPENAI MODEL SPEC OF OCTOBER 29, 2026: CONFIRMED FACTS AND OPEN QUESTIONS.

The established facts are the structure and content OpenAI has published. The October 29, 2026 edition organizes model behavior around faithfulness to human intent, harm minimization, honesty and transparency, human agency and care. It refers to the model as GPT, removes most detailed style rules, rewrites the chain of command, adds guidance on context, memory and social norms, makes corrigibility a principle of its own, and defines a scope of autonomy for agentic work with explicit authorization required for high-risk activities.

The open questions concern implementation. OpenAI has said its models do not yet fully reflect the document, has not published compliance measurements against the new text, and has described its value-based training methods only as work in progress. Whether removing detailed style rules leads to noticeably different ChatGPT responses, and how the scope of autonomy will be expressed in products such as Codex or in the API, will only become clear when OpenAI ships models trained against this version and reports how closely they follow it.

··········

·····

DATA STUDIOS

·····

datastudios.org