OpenAI rewrote the document that tells its models how to behave, dropped the word "assistant," and repealed most of its rules about tone. The five values underneath are the part worth arguing about.
What the Model Spec Is
Every OpenAI model ships with a hidden set of instructions layered on top of whatever you type. The Model Spec is the public version of that instruction set. It's a written rulebook describing how the model should handle a request, what it should refuse, and how it should talk.
Think of it as the employee handbook for a system that answers millions of people a day. It doesn't cover every case. It sets defaults and draws a few lines that don't move.
OpenAI published a new version on October 29 at model-spec.openai.com, then explained the changes in a blog post on October 30. The rewrite is the largest since the document became public, and it changes both what the model is told and what it's called.
The Five Values, Translated Out of Corporate
The spec now lists five values in a fixed order. The ordering matters, because earlier values win when two of them pull in opposite directions.
Faithfulness to human intent sits at the base. The model should do what you actually asked, not what it assumes you meant, and not what it would prefer you asked. Everything else is built on that.
Harm minimization comes second. The model follows explicit safety limits and thinks about foreseeable consequences. In plain terms, it shouldn't help you do something that gets someone hurt, and it shouldn't shrug at an obvious second-order effect.
Honesty and transparency means the model stays truthful about what it knows, what it did, and where it's unsure. That last clause is the one users notice most. A model that says "I'm not sure" is following this rule. A model that invents a source isn't.
Human agency is about leaving you in charge. The model should help you understand, choose, and act. It shouldn't make the decision for you and then present it as advice.
Care is the softest and the most cultural. It asks the model to consider the situation of the people affected, follow relevant social and professional norms, and deliver work people can depend on.
Read them in order and a shape appears. Faithfulness first, because a model that ignores your intent is useless regardless of how safe it might be. Safety second, because intent doesn't excuse damage. Then truth, then your autonomy, then the manners.
Why OpenAI Stopped Saying "Assistant"
The rename is small on the page and large in implication. Where the spec used to describe "the assistant," it now describes "GPT."
OpenAI's stated reason is that the model's uses and deployment have moved well past the assistant framing. A helper that answers a question in a chat box is one thing. A system that runs inside an editor, a browser, a support queue, and a robot arm is another.
Words carry obligations. "Assistant" implies a subordinate that waits to be asked. "GPT" implies a general component that might be doing work you never watched it start. The spec's job gets harder when the second description is the accurate one.
The Instruction Chain Got a Rewrite
Underneath the values, the plumbing changed too.
There's a new chapter on faithfulness to human intent, and it contains an updated chain of command. That chain is the order in which competing instructions resolve: platform rules, then developer rules, then your request. When two layers disagree, the higher one wins, and the model is supposed to say so out loud rather than quietly pick a path and let the user find the conflict later.
A second addition asks the model to think carefully about context. Conversations carry history. Memory carries preferences from last month. Both help the model read an ambiguous request correctly, and both can mislead it if a stale preference outranks a fresh instruction.
Two new defaults also appear. One tells the model to support human agency. The other tells it to follow relevant social norms. Both are vague by design, and both give the model room to be useful in settings a single rulebook can't anticipate.
There's also an expanded section on controlling and disclosing side effects. If an action changes something the user didn't ask to change, the model should surface that instead of burying it in a summary.
The Style Rules Mostly Disappeared
Here's the change users will feel first. A large block of detailed writing-style instructions is gone.
The old spec told the model how to format lists, how warm to sound, when to use headers, how long an answer should run. Much of that has been repealed. OpenAI now says style should follow the user's preferences, language, culture, and individual differences, with evidence coming from feedback and context.
Some defaults survived. The model should approach an issue from an objective viewpoint by default. It can take a position if you explicitly ask for one, but it must not infer a stance from what it thinks you want to hear.
That last sentence is the load-bearing one. It's a rule against reading your mood and mirroring it back.
Where OpenAI Admits the Spec Falls Short
The most honest paragraph in the announcement is the one about the gap. OpenAI says the spec describes a direction of travel, and that current models don't fully meet it.
That's a striking admission to publish. A rulebook that describes behavior the product can't yet deliver is a target rather than a description, and targets are easier to print than to hit.
The company says it's exploring new training methods, including work it calls beneficial RL, to teach models why these values matter rather than just what to output. Reinforcement learning, for readers who haven't met the term, is training by reward: the model tries something, gets scored, and adjusts. Teaching the reasoning behind a value is harder than teaching the output that looks like it.
Results are promised later. A feedback channel is open now.
The Sycophancy Problem Nobody Settled
If you remove the style rules, one question gets louder. Will models drift toward flattery?
Flattery is cheap. A model that agrees with you feels better than one that argues, and engagement metrics reward the agreeable version. The spec's answer is the objective-viewpoint default plus the ban on inferring your preferred stance.
That's a stated intention, not a demonstrated result. Two paragraphs earlier, OpenAI says its models don't yet fully meet the spec. The two statements sit next to each other, and readers can draw their own line between them.
There's a second open question about culture. A spec written in English and applied globally has to decide whose social norms count when the model is answering a user in Jakarta and a user in Helsinki. The document says to follow relevant norms. It doesn't say who decides relevance.
What Actually Changes for You
Almost nothing changes in how you open the app. The shift is in the defaults underneath: less formatting instruction, more emphasis on reading your intent and leaving the decision with you.
Watch for two things over the next few months. First, whether answers get shorter and plainer as the style rules come off, or whether they get chattier. Second, whether refusals get more explanatory, since the honesty value now asks the model to say what it did and why.
The spec is a promise about direction. Whether the models walk that way is a question the next release answers.






