OpenAI has rewritten the Model Spec, the public document describing how its models should behave, so that it rests on five values instead of a long list of rules. The company announced the revision on its own site on 29 October 2026. It also deleted most of the document’s detailed guidance on style.
Faithfulness to human intent leads the list and anchors the rest. The other four are care, honesty and transparency, harm minimization, and human agency. According to OpenAI’s announcement, it means working out what OpenAI, developers, and users each want, and acting sensibly when that picture is incomplete. The chain of command, which settles whose instructions win in a conflict, now lives inside that section.
OpenAI offers a small example. A model helping someone compare job offers might draw on a priority the person mentioned earlier, such as predictable hours, and say that it is doing so. The user can then correct the assumption. Memory helps in this picture only when the person can see it and overrule it.
The revision adds a subsection on reading context, covering how conversations, memories, and other evidence should shape the model’s guess about what someone wants. It also adds defaults for backing human agency and for respecting social norms, and expands the guidance on side effects of the model’s actions and how to tell users about them.
The retired material is the part developers will notice. OpenAI now says style should follow what people prefer, shaped by language, culture, and the individual, with user feedback as the evidence. A few firmer defaults survive. One is that the model approaches questions objectively and does not adopt a viewpoint just because it suspects the user wants to hear it. Users and developers can still request a perspective explicitly. One cosmetic change: the Spec now calls the model “GPT” where it used to say “the assistant.”
The logic is easy to follow. A model that books travel, edits files, or runs a workflow meets situations no instruction anticipated, and a rulebook cannot list them all. Companies handle the same problem with people by writing down what the job is for and trusting judgment at the edges. The trade is predictability. A fixed style rule behaves the same way every time, while an inferred preference varies by user, language, and context.
OpenAI is candid about the gap. The post says the Spec describes the behavior the company is working toward and that its current models fall short of it. It mentions new training methods meant to teach the models why the principles matter, and promises results in later updates. None have been published, and the announcement includes no outside evaluation of how closely current models follow the five values.
That leaves a question the document cannot answer yet: how anyone, inside or outside OpenAI, would measure faithfulness or care. Values are harder to test than rules. Until a test exists, the Spec reads as a statement of intent from the company that writes it.
Teams that tuned prompts around GPT’s old default tone and formatting should re-run those prompts against current behavior, because the style rules they relied on are no longer written down.
Published by OpenAI on 29 October 2026.
