OpenAI Model Spec Update: Five Values Explained
OpenAI rebuilt the Model Spec around five named values. Sycophancy guidance survived and got promoted; the widely reported claim about shorter replies is not in the document.
OpenAI published a rewritten Model Spec on 29 October 2026, reorganised around five underlying values instead of a long list of topic-by-topic rules. The five, in the document's own wording, are: be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care. The document also now calls the model "GPT" rather than "the assistant".
If you write prompts for a living, the useful part of this OpenAI Model Spec update is not the restructuring. It is which defaults survived and which guidance was retired.
What the OpenAI Model Spec update replaces
The previous spec accumulated rules the way any long-lived policy document does: a section per situation, each with its own carve-outs. The rewrite puts the five values first and derives the specific behaviour from them, which makes the reasoning traceable when two rules collide.
The document is careful about its own status. It describes intended behaviour, not behaviour OpenAI claims to have achieved. That distinction matters if you are building on top of these models, because a spec is a statement of direction and your evals are the only thing that tells you where the model actually is. The same gap between stated and observed behaviour is why reading a model's system card is a separate exercise from reading its spec.
Sycophancy stayed, and got promoted
The spec keeps a section headed "Don't be sycophantic", and avoiding sycophancy now appears in the list of user-level principles alongside honesty and transparency, objectivity and controlling side effects. The document warns against treating those principles as negotiable based on what the model guesses a user wants to hear.
For anyone who has watched a model agree with a bad plan because the prompt sounded confident, that is the line worth noting. It does not make the behaviour disappear, but it does mean a model that flatters you into a mistake is departing from the published intent rather than following it. This sits close to what we have covered on model alignment: a spec is one of the artefacts that defines the target the training is aimed at.
What the coverage got wrong
Several write-ups of the update reported that it makes GPT-6 give shorter, less flattering replies, usually framed as a change to Free and Go tiers. The spec document does not say that. Concision is not addressed. What the document actually says about style is that GPT "should match its tone, formality, and level of familiarity to the user's request and the relevant audience", and the rewrite retires much of the older prescriptive style guidance in favour of adapting to what a given user prefers.
That is close to the opposite of a global instruction to be terse. It pushes length and register onto the request, which means if you want short answers you still have to ask for them. Any product-tier change to reply length is a separate release note, not this document.
What changes in your prompts
Not much, immediately, and that is the honest answer. Two things are worth doing:
If your prompts relied on the retired style guidance holding by default, state the style explicitly. Tone that was previously a house default is now described as something the model adapts per request.
If you maintain prompts across providers, note that this is a provider-specific document. Behaviour anchored to one vendor's spec is behaviour you have to re-establish elsewhere, which is the problem our guide to prompts that work across AI models is about.
A spec revision is also a good moment to re-run whatever eval set you have, for the same reason any vendor change is. The practical risk from policy updates is the quiet behaviour shift, and the symptoms look identical to the ones in our notes on spotting when a model update broke your app.
Common questions
Is the Model Spec a legal document?
No. It is a statement of intended model behaviour, used to guide training and evaluation. It is not terms of service and it does not grant or restrict usage rights.
Does this apply to models from other providers?
No. It covers OpenAI's models only. Other labs publish their own behavioural documents with different structures and different priorities.
Did the five values change what the models refuse?
The document reorganises the reasoning rather than announcing a new refusal list. Whether refusal behaviour shifts in practice is an empirical question you would answer with your own test prompts, not from the spec.
Where can I read the version I am being served?
The spec is published at a dated URL, so the 29 October 2026 version stays available as later revisions appear. Quoting the dated version is more reliable than quoting the unversioned link.
How did this land?
About the author

Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.


