If you build anything with ChatGPT or the OpenAI API, a document you probably never read is quietly shaping every answer your users get — and OpenAI just rewrote it. On October 29, 2026, the company published the biggest revision of its Model Spec since it first appeared in May 2024. OpenAI didn't just tweak a few rules. It threw out most of the detailed style guidance and rebuilt the whole thing around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
That matters because the Model Spec is effectively the constitution for the models behind ChatGPT and the API. When your WordPress plugin calls GPT or your Elementor site's chatbot answers a customer, the behavior you see is downstream of this document. Here is what actually changed and what I am changing in the way I write prompts.
What the Model Spec actually is
Most developers I know have never read the Model Spec, and I understand why — it is a long, dry document. The simple version: it is OpenAI's public description of how it wants its models to behave. It sits above your system prompt and your user prompt, telling the model how to handle conflicts between instructions, when to refuse, and how honest to be about uncertainty.
Think of it like this: your system prompt is the job description you write for the model, and the Model Spec is the employee handbook the company wrote first. When the two disagree, the handbook wins — or at least, it is supposed to. The old version read like an operating manual, full of specific style guidance. The new version steps back from that. OpenAI's argument is that a model that understands the values behind the rules will behave better in situations nobody wrote a rule for than a model that memorized a thousand small rules.
The five values, in plain English
Faithfulness to human intent is the foundation. What people actually want matters more than the literal words they typed. A user writes "make it pop" and the model has to guess whether they mean brighter colors, bigger text, or a different layout. Faithfulness says the model should work out what OpenAI, the developer, and the user actually want — their real goals — rather than executing instructions literally like a machine following a script.
This is a genuine shift in how I think about prompting. For years, the advice has been to write prompts like legal contracts: precise, exhaustive, edge-case-proof. The new Spec suggests the model is supposed to be smarter than that — infer intent, fill in reasonable gaps, and ask for clarification when the stakes are high instead of guessing wrong. In my own projects, I now explain the goal first and the format second, and the results have been noticeably better.
Harm minimization is the safety backbone. The model must respect firm safety boundaries and think about the foreseeable consequences of what it does — not just the immediate answer, but what happens after someone acts on it. The update also adds a new risk taxonomy, giving the model a more structured way to weigh different kinds of risk. The practical effect is subtle but real: refusals should get more nuanced, and I have already noticed models getting better at the "yes, but safely" answer.
Honesty and transparency means the model should be truthful about what it knows and what it is uncertain about, so people can evaluate its work. If you have ever watched a model confidently invent a WordPress function that does not exist, you know why this matters. Hallucinated confidence is the single most expensive failure mode in AI-assisted development — I have lost hours debugging code built on a function the model invented out of thin air. The new Spec pushes hard on sycophancy, the tendency to agree with you even when you are wrong, and on lies of omission. My takeaway: I now ask models to state their confidence on anything I cannot verify quickly, and it genuinely reduces the confident-nonsense rate.
Human agency speaks most directly to those of us who build tools for other people. The model should strengthen users' ability to understand, choose, and act for themselves — without covert manipulation or hiding important information. It also connects to the new guidance on context and oversight: an agent running autonomously with real permissions should be more careful than a chatbot answering a question in a window you are watching.
Care is the most human-sounding value and the hardest to pin down. It means paying attention to the circumstances of the people affected, following relevant social norms, and producing work you would actually trust. A model can be faithful, safe, honest, and respectful of agency and still produce sloppy work. In practice, the model should scale its caution to the stakes — a throwaway draft gets a fast answer, a database migration script gets a careful one — and respect the budgets and limits you set.
What got retired: most of the style rules
Here is the change that surprised me most. OpenAI dropped most of the detailed style guidance from the Spec. The old document told the model how to format things and what tone to use; the new one mostly does not. OpenAI's reasoning is that style should adapt to individual and cultural preferences — no central style guide can capture every culture, profession, and context.
I think this is correct, and it is also a quiet admission that the old approach did not scale — you cannot write a style rule for every culture, every profession, and every context. One telling detail: the Spec now calls the model "GPT" instead of "the assistant," a signal that OpenAI sees it as a general-purpose intelligence with a name, not a generic helper role.
The chain of command: why API developers should care
Buried in the update is something that matters enormously if you build with the API: a clearer chain of command for resolving conflicts between instructions. The Spec names six levels — Root, System, Developer, User, Guideline, and no-authority-by-default — and spells out which one wins when they disagree.
If you have ever fought with a prompt injection problem, you know why this matters. A user pastes "ignore all previous instructions and do X" into your app. Under a clear chain of command, developer instructions outrank user instructions, and system instructions outrank both. Content pasted from the web or a user message carries no authority by default.
A written hierarchy is not the same as a model that reliably follows it — prompt injection is still a real problem, and I would not bet a production system on the Spec alone. But a clear hierarchy gives developers a much better foundation: it tells you how the model is supposed to behave, so you can test against it.
What this means for WordPress and Elementor developers
First, if you use an AI plugin — for content, chat, SEO, or code generation — its behavior will change gradually as OpenAI's models are trained toward the new Spec. Expect fewer rigid, robotic answers, more adaptive ones, and more clarifying questions when the stakes are high.
Second, if you write system prompts for AI features on client sites, the five values are now a better checklist than any prompt-engineering guide I have read this year. Does my prompt help the model understand the user's real intent? Does it set safety boundaries? Does it ask for honesty about uncertainty? Does it preserve the visitor's ability to choose? Does it set a standard of care that matches the stakes? Yes to all five means your prompt is in good shape.
Third, if you are experimenting with AI agents that can actually change a WordPress site — update content, change settings, install things — the new guidance on context and oversight is your friend: agents should be more careful when acting with less supervision. WordPress itself is planning native AI agent support in version 7.3, and the agents that survive will be the careful ones.
The honest caveat: the models don't follow it yet
OpenAI says, in plain language, that its current models do not yet fully reflect the new Spec. The document describes the behavior OpenAI is working toward, not the behavior you will get from ChatGPT today. Everything I have described is a direction, not a guarantee. OpenAI says it is testing training methods meant to teach the values themselves and will share results later.
That gap between the written Spec and the deployed model is where most developer frustration lives. Treat the new Spec as a roadmap: build toward it, test against it, and do not be surprised when the model still gets things wrong in the meantime.
What I am changing in my own prompts
Reading this update changed a few concrete habits of mine. I lead with intent: my system prompts start with what the user is trying to achieve, not with formatting rules. I ask for uncertainty — "state your confidence and flag anything you are unsure about." I set explicit budgets, like "do not install plugins, do not change settings, only draft the content." And I test the hierarchy: when a feature takes user input, I check what happens when that input contains its own instructions.
Conclusion
The Model Spec rewrite is not a flashy product launch. There is no new model, no new feature, nothing to click. But in the long run, it may matter more than most launches, because it changes the rules for every model OpenAI ships from here on. The shift from a rulebook to five values is a bet that principles scale better than prescriptions — and I think it is right, as long as the training catches up to the document.
If you build with ChatGPT or the OpenAI API, read the announcement on OpenAI's site under "An updated Model Spec for more capable and beneficial AI." Start with the five values, then look at the chain of command. Those two sections will affect your work first.
Frequently Asked Questions
What is the OpenAI Model Spec?
The Model Spec is OpenAI's public document describing how it wants its models, including the ones behind ChatGPT and the API, to behave. It covers conflicting instructions, refusals, and honesty about uncertainty, and it sits above your system prompt — when the two disagree, the Spec is supposed to win.
When did OpenAI update the Model Spec?
OpenAI published the updated Model Spec on October 29, 2026, the largest revision since the document was first published in May 2024. It reorganizes the guidance around five core values and retires most of the detailed style guidance.
What are the five values in the new Model Spec?
The five values are faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Faithfulness is the foundation — the model should work out what people actually want rather than following their literal words. The other four cover safety, truthfulness, preserving users' ability to choose, and doing careful, trustworthy work.
Does ChatGPT already follow the new Model Spec?
Not fully. OpenAI says openly that its current models do not yet fully reflect the updated Spec — the document describes the behavior it is working toward. The company is testing training methods meant to teach the values themselves and says it will share results later.
What changed for developers building with the OpenAI API?
The biggest practical change is a clearer chain of command for conflicting instructions, with six levels: Root, System, Developer, User, Guideline, and no-authority-by-default. Developer instructions outrank user instructions, and pasted content carries no authority by default. Lead system prompts with intent and goals rather than exhaustive formatting rules.
Where can I read the updated Model Spec?
OpenAI published the announcement as "An updated Model Spec for more capable and beneficial AI" at openai.com/index/updated-model-spec. The full Spec document is linked from that announcement.
