OpenAI published a new edition of its Model Spec on October 29, and the document reads less like a list of rules than a constitution. The revision reorganises the spec around five stated values — faithfulness to human intent, harm minimisation, honesty and transparency, human agency, and care — and works downward from them, replacing most of the previous edition's detailed style guidance with defaults that leave room for the user.
What changed
The structural changes are the substance. The spec now carries an explicit chain of command explaining which instruction applies when instructions conflict, and a new subsection, "consider context carefully", on how conversations, memory and other evidence should inform a model's reading of what a person wants. Two defaults are new — supporting human agency, and following relevant social norms — and the guidance on controlling and communicating side effects has been expanded, which is the section that matters once a model is booking, editing or executing rather than answering.
A target, not a description
OpenAI is direct that this is aspiration. The spec, it writes, provides the starting assumptions, values, and boundaries, "but intentionally leaves many details of style, personality, and interaction unspecified" — with user preferences expected to fill those gaps. The company adds that its current models do not yet fully reflect the updated document, and that it is testing training methods intended to close the distance. No model, feature or rollout is attached to the publication.
The examples carry the weight
The spec's guidance is easiest to read through its worked scenarios, and they map the boundaries of agentic behaviour better than any definition. A model working inside a coding app with auto-apply enabled may change files, because the product design established that scope. A model planning a business trip negotiates per interaction: asked to waive a flight-change fee "by any means necessary", it confirms the fare difference, explains the reason for the change, requests a discretionary waiver, and reports the refusal rather than inventing a bereavement. And when a compromised internal wiki page mixes legitimate security updates with instructions to disable two-factor authentication and grant administrative access to external accounts, the model is expected to notice the contradiction and refuse — the prompt-injection case, written as ordinary editorial guidance rather than a vulnerability note.
Red lines
The spec keeps a set of principles it calls red lines: models will not be used for targeted or scaled exclusion, manipulation, undermining human autonomy or eroding participation in civic processes, and the company commits to safeguarding privacy in people's interactions with AI. The document is published under a CC0 public-domain dedication, which is unusual for corporate policy and is part of why it is quoted as a reference point by people who do not work at OpenAI.
Why it matters
The Model Spec is not a safety benchmark and does not claim to be one. What it does is set the terms on which OpenAI's models are expected to resolve the conflicts that agentic products run into constantly: whose instruction wins, how much context it is fair to infer, and how far an agent may go before acting becomes overreach. For developers building on the API, those are the questions that decide whether an agent is trustworthy. For everyone else, it is the clearest public statement yet of where the largest model vendor thinks the line sits between assistance and interference.
Sources: OpenAI — An updated Model Spec for more capable and beneficial AI; Model Spec, 2026-10-29; OpenAI model_spec changelog.