OpenAI rewrites the Model Spec around five values, retires most of its style guidance, and starts calling the model GPT instead of the assistant

The source. The primary material is OpenAI's own post, An updated Model Spec for more capable and beneficial AI, published on openai.com on October 29, 2026, together with the new version of the document itself, v2026.10.29, which carries that date in its footer. The CHANGELOG in the openai/model_spec GitHub repository lists the same version as its most recent entry. Every addition, removal and rename below is taken from those two places and the changelog, not from secondary summaries, and nothing the material does not state has been filled in.

Five values. The document goes from a long list of parallel behaviour rules to five values with implementation detail underneath each. The five, in OpenAI's words: faithfulness to human intent, minimising harm, honesty and transparency, facilitating human agency, and acting with care. OpenAI's own explanation is that these commitments were already scattered through the old version, just never stated up front, so pulling them together makes the reasoning behind each rule legible. Faithfulness to human intent sits at the top, which means the entry point to the whole document is now understanding what a person is trying to do rather than executing their words literally. That sentence is the real through-line of the update.

The new Consider context carefully section is the most practical part of the update. It sets seven behaviours around context: carry relevant context forward, use it proportionately, weigh source and scope and recency, respect privacy and audience, preserve user control, and make context use legible and steerable. In plain terms: memory and personalisation are allowed, but they have to be restrained, visible and easy to correct. OpenAI's example is concrete. When helping someone compare job offers, the model may draw on priorities that person shared earlier, such as wanting predictable hours, but it has to surface that assumption so it can be corrected at a glance. This section is really a behaviour spec bolted onto memory. In the 2025 versions memory was a product feature; now there is a rulebook for when using it counts as overreach.

Three additions to the chain of command. The Root, System, Developer, User and Guideline tiers were already there; this version adds support human oversight and correction, assume best intentions, and ignore untrusted data by default. The third looks the most technical and matters the most in practice: web pages, retrieved documents and output from third-party tools all count as untrusted by default and must not be executed as instructions. Anyone who has built an agent will recognise why that line exists, because almost every prompt-injection entry point lives there. Assume best intentions adds slack on the other side, telling the model to read ambiguous wording as a reasonable request rather than defaulting to a defensive reading of the worst case.

Control and communicate side effects moves up from a guideline to a user-level principle, with new language on authorisation and trade-offs. This one governs actions with real consequences: deleting files, sending mail, placing orders, changing production configuration. User-level means it outranks ordinary style advice, so a model cannot skip a confirmation round-trip just to look helpful. The two other new defaults land in similar territory. Supporting human agency asks the model to strengthen a person's own ability to judge and act rather than deciding for them, and following relevant social norms asks it to remember that the work will be seen by other people, so write with the reader in mind.

The deletion is the story. OpenAI states plainly that it is retiring most of the Spec's detailed style guidance, and the reasoning is just as plain: style should adapt to what people prefer, in aggregate, within languages and cultures, and for each individual, with user feedback and context providing the evidence. That is hard to argue with, and it also means the lab has given up its preset for tone. Only a few strong defaults survive, and the one OpenAI names is approaching questions objectively: users and developers may explicitly ask the model to take a position, but it should not infer that preference from a guess about what they want to hear. Compare the September 12, 2025 version, which pushed in a fuller set of default personality principles and a section on using appropriate style. The direction has reversed.

Read the last three versions together and the line is clear. On September 12, 2025 the top authority tier was renamed from Platform to Root with Root above System, so the highest principles cannot be overridden by a system message, and principles for agents that act in the world were added. On August 18, 2026 the spec filled in teen relational guidance, added a section on being clear about capabilities and limits, and dropped outdated advice about pre-reasoning models. On October 29, 2026 it reorganised around values, retired style guidance and changed the name. What all three share is a push from what a model should do toward what a model should understand, from a rule list to a judgment framework. That matches the motive stated in the post: as models take more consequential actions, reading instructions literally is no longer enough.

The rename looks like housekeeping and is actually the whole logic in miniature. The assistant presumes a position: passive, waiting to be asked. GPT is just a model name and presumes nothing. OpenAI's stated reason is that the range of use cases and deployment settings has widened past what one word can hold. That is politely put, but the implication is that the model is no longer always the thing that moves only when called. In some settings it executes, in others it reviews, in others it is an agent running for hours. Change the name and the responsibility boundary has to move with it, which is why support human oversight and correction appears in the same version's chain of command.

How to read it. First, this describes intended behaviour, not current behaviour. OpenAI says outright that its production models do not yet fully reflect the Model Spec, so treating the document as a description of what ChatGPT does today is a misreading. Second, the five values are not hard constraints. OpenAI says that beyond the safety red lines most of the content is defaults that users and developers can override, which means the thing that actually governs your product is your own system message, not this public document. Third, and most useful: the changes themselves are the signal. Retiring the style rules, writing oversight and correction into the chain of command, and treating untrusted data as untrusted by default all point at one judgement — in late 2026 OpenAI is less worried about models that cannot do the work than about models that do the wrong work and argue convincingly that it was right.