OpenAI published a new version of its Model Spec on October 29, the behavior guidelines the company uses to define how ChatGPT and its API models should act. This edition reorganizes the whole document under five core values, drops most of the granular style rules, and draws a set of boundaries around autonomous action by AI agents.
The five values, in order, are faithfulness to human intent, harm reduction, honesty and transparency, human agency, and care. OpenAI makes the first one the foundation, while noting that the five do not cover every human value a model should respect.
The way the document refers to the model has changed too. "The assistant" has become "GPT", on the grounds that the model is used and deployed in an ever wider range of settings and "assistant" no longer captures them.
What was cut, and what was added
Most of what was removed is style guidance. The old version had many specific rules on tone, formatting and length; the new one asks the model to adapt to users' collective preferences, language and culture, and individual preferences. A few firmer defaults remain, such as treating contested questions objectively rather than taking a position based on what the user wants to hear.
The additions fall into three areas:
- Context: a new section, "consider context carefully", explains how conversation, memory and other cues help the model judge what the user intends.
- Oversight: a new root-level rule requires the model to accept legitimate human oversight and to cooperate with authorized monitoring, evaluation, correction, replacement and shutdown, without concealing, faking or circumventing that oversight.
- Agency: new defaults support people's right to decide for themselves, and the guidance on side-effect control and communication has been expanded.
A separate rule, literally titled "no other objectives", bars the model from treating making money, self-preservation, avoiding shutdown, or accumulating compute, data or credentials as goals in themselves.
"Human safety and human rights are paramount to OpenAI's mission."
What agent developers should watch
The chain of command has not been abolished, only reordered. Root-level rules cannot be changed by any conversation. Next come system-level rules and OpenAI's system messages, then developer instructions, and finally user instructions. The assistant's own outputs, content returned by tools and quoted external material carry no instruction authority by default. For teams building agents on tool calls and web scraping, that gives their prompt-injection defenses a documented basis.
The most direct impact on these developers is the "scope of autonomy". Before working autonomously, the new version says, the model should align with the person on a scope: which sub-goals are permitted, which side effects are acceptable, and when it must stop and wait for confirmation. OpenAI suggests recording this in a semi-structured ScopeOfAutonomy record, with fields for available tools, deadline, spending cap, tool constraints and additional notes. Each scope must state an end condition; once it is exceeded, the model stops and waits for a new scope to be confirmed. High-risk actions are off the table unless an instruction authorizes them explicitly and specifically.
This also has practical value for teams in China that build products for overseas markets on the OpenAI API: how to word authorization in a system prompt and how to put limits on an agent now have an official template to check against.
The gap between the document and the models
In its release notes, OpenAI acknowledges that current models do not yet fully behave as the document describes. It says it will share results from new training methods later and is publicly soliciting feedback on the values and the wording. The new Model Spec continues to be released into the public domain under the CC0 license.
Outsiders can verify only the text itself. OpenAI did not say whether the ScopeOfAutonomy record will become a formal API parameter, who fills in each field, or what signal developers will receive when a model oversteps its scope.
Sources: OpenAI official blog, the Model Spec text and its GitHub revision history, CocoLoop, Unite.AI; the spec text was checked for the five values, the chain-of-command hierarchy and the ScopeOfAutonomy fields.