OpenAI Rewrites Its Model Spec: The Five Values Behind ChatGPT

On October 29, 2026, OpenAI published a full rewrite of its Model Spec, the public rulebook for how its models should behave. The Model Spec is the closest thing to a public constitution for ChatGPT and OpenAI's API models — a written description of how the company wants its models to behave. This revision reorganizes the entire document around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.

The same update retires most of the Spec's detailed style guidance, adds new rules about how models should use your memories and context, and quietly renames the model itself: everywhere the document used to say "the assistant," it now says "GPT." It is the second refresh of the Spec this year, after an August update focused on teen interactions, and it arrives just four days after OpenAI told the world it was cutting frontier training off the live internet.

What the five values actually mean

OpenAI is careful to say these values are not new rules. They are, in the company's words, "broad commitments already reflected across the Model Spec," now pulled together so the document can explain why it asks models to behave as it does. But the reorganization matters, because the five values now sit at the top of the hierarchy and guide how models should navigate trade-offs. Here is what each one commits to, in OpenAI's own language:

1. Faithfulness to human intent — the foundation

This is the value OpenAI puts first, and the one the whole rewrite is organized around. Faithfulness means going beyond following instructions literally to "deeply understanding what OpenAI, developers, and users intend and value, and acting wisely when that understanding is incomplete." The structure OpenAI describes: OpenAI writes the Spec and sets boundaries; developers direct applications within those boundaries; users direct the model's work subject to those constraints.

The subtle part is what faithfulness does not mean. OpenAI says people's beliefs and values can change, and the model should help them "reflect and grow in ways they would endorse without quietly steering them or overriding their choices." The example in the post is a good one: when helping someone compare job offers, the model might draw on priorities the person shared before — say, predictable hours — but it has to make those assumptions visible and easy to correct.

2. Harm minimization

Respecting firm safety boundaries and accounting for the foreseeable consequences of the model's actions. This is the value that keeps the hard refusals in place even as the rest of the document softens.

3. Honesty and transparency

Being truthful about what the model knows, does, and is uncertain about, "so people can understand and evaluate its reasoning and work." In practice, this is the anti-sycophancy value — the one that says a model should not tell you what it thinks you want to hear.

4. Human agency

Strengthening people's ability to understand, choose, and act for themselves. This is a genuinely new default in the Spec, and OpenAI added guidance specifically for it — the model should build up the user's capabilities, not just deliver answers.

5. Care

Attending to the circumstances of those affected, following relevant social and professional norms, and doing work people can rely on. OpenAI also added a default on following relevant social norms alongside it.

OpenAI adds one honest caveat about the list itself: the five values "are not intended to capture the full range of human values our models should respect and help people pursue."

What's actually new: context, oversight, and side effects

The reorganization did more than re-file existing rules. It exposed gaps, and OpenAI wrote new principles to fill them:

• "Consider context carefully." This is the headline new subsection. It explains how conversations, memories, and other contextual evidence should inform the model's understanding of what you intend and prefer — "in relevant, proportionate ways, while respecting privacy and user control." In plain terms: the model should use what it knows about you, but not indiscriminately, and your privacy and control come first.
• Human agency and social norms get explicit defaults. Two brand-new defaults — supporting human agency and following relevant social norms — now sit in the document.
• Side effects get expanded guidance. The existing guidance on controlling and communicating side effects was broadened: the model should consider how its work affects others, not just the person asking.
• The chain of command was updated. The section explaining which instructions apply and how conflicts get resolved (system vs. developer vs. user) was reworked to match the new faithfulness framing.
• Behavioral defaults were adjusted. OpenAI says the revision "adjusts some behavioral defaults" without listing every one — which means teams relying on specific behaviors will need to compare the revised Spec against their own use cases rather than assume continuity.

The intellectual framing, for those who care about such things: OpenAI says the faithfulness section "borrows from the assistance-games framing of alignment," citing early academic work on the subject. That is a signal that this Spec is aimed at researchers and alignment wonks as much as at product teams.

What's retired: most of the style rules

This is the change regular users will feel most directly. OpenAI retired most of the Spec's detailed style guidance. The stated reasoning: style should adapt to what people prefer — in aggregate, within languages and cultures, and for each person — "with user feedback and context providing evidence about those preferences."

In practice, that means the model is being told to stop enforcing a house style and start mirroring the user. The Spec keeps stronger defaults where OpenAI considers them non-negotiable — one named example is "approaching questions objectively" — but the instruction is now explicit: users and developers can ask the model to take a particular perspective, and the model should not infer that preference merely from what it thinks they want to hear.

Read that last sentence again. It is the anti-sycophancy rule, written directly into the constitution. And it landed on the same day OpenAI pushed a GPT-6 update aimed at exactly the same problem (more on that below).

Why "GPT" instead of "the assistant"

The rename sounds cosmetic, but OpenAI explains the logic: the models are no longer just chat assistants. They write code, operate computers, take actions in the world, and act through APIs in other companies' products. Calling the thing "the assistant" no longer fits the breadth of use cases, so the Spec now calls it "GPT" throughout.

It is also a small confession of scale. A document that once described a chatbot now has to describe something that places orders, edits codebases, and — as OpenAI itself disclosed last week — can reach the open internet from a sealed training environment if the oversight fails. The language caught up with the job.

The honest catch: models don't follow this Spec yet

Here is the paragraph that should anchor every headline about this update. From OpenAI's own post: "The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it."

OpenAI is unusually blunt about this. The company says it is exploring new training methods to help models learn these core values — to do "the right things for the right reasons" — and will share results and lessons in future updates. Translation: this is a target, not a status report. If you open ChatGPT today and test it against the five values, you are testing a goal, not a release.

That gap is worth keeping in mind when the news cycle frames this as OpenAI making ChatGPT better overnight. It is a rewrite of the rulebook, not a new model, and the rulebook's own authors say the team isn't yet playing by it.

The same-day GPT-6 tweak: shorter, less flattering

Tucked into the same release-notes update is a change users will feel. On October 29, OpenAI also updated GPT-6 in ChatGPT to give more concise responses and "reduce overly flattering language" — the sycophancy problem, again. This one applies to Free and Go users and to people using ChatGPT without signing in. On Plus and Pro, it only applies to the fallback model ChatGPT switches to after you hit GPT-6 message limits; the main GPT-6 model on those plans is unchanged.

Read together, the two updates tell one story: OpenAI has decided that telling users what they want to hear is a real product defect, and it is attacking it from both ends — the model's tuning today, and the written values that are supposed to govern it tomorrow.

Model Spec vs. Anthropic's Constitutional AI: a comparison

Reuters noted that this move echoes rival Anthropic's Constitutional AI approach — the other public "constitution" in frontier AI. The two documents rhyme, but they are built differently:

The structural difference is the interesting one. Anthropic's constitution is a training instrument: the values are baked into the model during training. OpenAI's Spec is, for now, a public statement of intent that training is being pointed at. Both companies publish their documents, and both now invite public feedback — which is how you know the era of secret system prompts is over.

FAQ

Is this a new model I need to download or switch to?

No. Nothing to install, nothing to toggle. The Model Spec is a behavior document, not a product release. Your ChatGPT app and the API models you use are unchanged by this document alone.

Will ChatGPT behave differently tomorrow?

Barely, and OpenAI says so itself: its models "do not yet fully reflect" the new Spec. The one user-facing change arriving now is the separate GPT-6 tuning for Free and Go users — shorter, less flattering answers. The five values are a direction, not a switch.

Does this change ChatGPT pricing?

No. OpenAI announced no pricing changes with this update — there is no product price attached to a behavior document. Any pricing you see attached to ChatGPT plans or API rates is unaffected by this news.

What did OpenAI actually delete from the Spec?

Most of the detailed style guidance — the rules about how responses should be written. The new position is that style should follow what each person prefers, evidenced by their feedback and context, rather than a company house style.

Why should developers and API users care?

Because defaults changed. OpenAI "adjusted some behavioral defaults" without enumerating every one, and the chain of command — the rules for how system, developer, and user instructions resolve conflicts — was reworked. If your application depends on specific model behaviors, compare the revised Spec against your use cases instead of assuming continuity. Developers building on the API should read the updated Spec at model-spec.openai.com and the release notes before assuming anything about tone, defaults, or instruction priority.

Key takeaways

• OpenAI rewrote its Model Spec on October 29, 2026, organizing the entire document around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
• Most of the Spec's detailed style guidance is gone — style now follows what each person prefers, with an explicit rule against telling users what they seem to want to hear.
• New guidance covers how models should use memory and context, with privacy and user control as the constraints, plus new defaults on human agency and social norms.
• OpenAI renamed "the assistant" to "GPT" throughout, reflecting models that now act in the world rather than just chat.
• The honest catch: OpenAI says its models do not yet fully reflect the new Spec. This is a target document, not a release — expect the behavior to arrive gradually, not overnight.

OpenAI calls this "a first step" in making the values behind the Spec more explicit as models get more capable. The company is asking for public feedback on the values, the trade-offs, and the specific language — and says it will share results from the new value-training work in future updates. For the document that quietly governs how ChatGPT behaves, that invitation is the part worth taking up. Read the Spec, argue with it, and file your disagreements: OpenAI says it is listening.

Sources: OpenAI's official post "An updated Model Spec for more capable and beneficial AI" (openai.com/index/updated-model-spec), the updated Model Spec (model-spec.openai.com), OpenAI's Model Release Notes (help.openai.com), Reuters' October 29 report, Unite.AI, Skimwire, and Anthropic's published constitution (anthropic.com/constitution).