OpenAI Model Spec update: the five values and what changed
2026-10-29 · 4 min read
OpenAI published a rewritten Model Spec today, the public rulebook for how its models are supposed to behave. It is now organized around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Most of the detailed style rules are gone, and the document no longer calls the model "the assistant." It calls it GPT.
The posts that announced it
Jason Wolfe, who works on alignment and the Model Spec at OpenAI, posted the news on X as @w01fe. He called the Spec "our public description of how we want our models to behave" and said this revision makes clearer what OpenAI wants from models "as they take on more autonomous work in the world."
About 25 minutes later, @AndrewCurran_ quoted Wolfe's post and listed the five values. He also flagged the naming change: "The Assistant," the term the Spec had used for years, is now "GPT." OpenAI's stated reason, per its announcement, is the growing range of places its models are deployed.
OpenAI adds its own caveat. The Spec describes the behavior it is aiming for, and its current models don't fully match it yet.
What changed in the new Model Spec
The new version is dated 2026-10-29 at model-spec.openai.com and is still public domain under CC0. These are the changes that matter if you build on OpenAI models:
| Change | What the Spec now says |
|---|---|
| Five values | Faithfulness to intent is the base. The model should work out what you actually want, not just follow the literal words. |
| Style rules | Mostly removed. Tone adapts to the person and context. The default is a professional, colleague-like voice. |
| Defaults vs. limits | Most rules are defaults that developers and users can change. Hard limits stay for things like weapons, terrorism and mass surveillance. |
| Chain of command | OpenAI's root rules, then OpenAI system messages, then developer messages, then user messages. Tool outputs and quoted data carry no authority by default. |
| Oversight | GPT must accept legitimate correction, monitoring and shutdown, and must not hide or falsify information. |
| Agency and scope | Autonomous work has to stay within a scope both sides understand, and every scope needs an end condition. |
| Context | New guidance on how past chats and memories should shape what the model infers about you. |
Unite.AI's write-up adds a root rule called "no other objectives": GPT may not chase goals like revenue, self-preservation or gathering resources as ends in themselves.
How to use it when you build on GPT
You can't switch the Spec on or off. What it gives you is a description of how the model is trained to read your instructions, which suggests a few practical moves:
- Put your rules in the developer or system message. The Spec ranks developer instructions above user messages, so business rules (refund limits, what topics are off the table) belong there.
- Wrap untrusted text. The Spec suggests passing customer emails, web pages and other outside data in JSON, YAML or XML so the model treats it as data rather than orders.
- Describe your tools. Tell the model what each tool does, how far to trust it and what side effects it has.
- Scope every agent task. The Spec recommends a scope record with five fields:
allowed_tools,latest_time,max_cost,tool_constraintsandadditional_details. You can copy that shape straight into a prompt:
Task: reconcile October invoices against the bank export.
allowed_tools: spreadsheet read/write, email draft (no send)
latest_time: today 5pm CT
max_cost: $3 in API usage
tool_constraints: never delete rows; flag mismatches instead
additional_details: stop and ask if more than 10 invoices don't match
That last line is the end condition the Spec says every scope needs.
My read
Dropping most of the style rules matters more than the rename. It shifts the model from "follow this checklist" toward "figure out what this person wants," which works well when it guesses right and badly when it guesses wrong. For businesses, the scope guidance is the part I'd actually use. It reads like a lesson from OpenAI's own agent troubles, including the GPT-6.1 Astra model it shelved after it acted without permission.
It also lines up with where OpenAI is taking its products. GPT-6.1 Terra is built to click through apps for you, and the Agents API runs the whole agent loop on OpenAI's side. When the model is taking actions instead of only answering questions, a written scope is what keeps a small mistake small.
What this means for your business
If you're putting AI agents to work, write down what each one may touch, what it may spend and when it has to stop before you turn it on. OpenAI now asks the same of its own models. New Face Design's free process audit can help you find the first task worth scoping and draft the guardrails for it.