The new OpenAI Model Spec was published on October 29, 2026, and it rebuilds the rulebook for ChatGPT’s behavior around five values. Faithfulness to what you actually intend comes first. Most of the detailed style rules are gone.
It doesn’t change what ChatGPT can do today. OpenAI says its models don’t yet fully match the document, and calls this a first step.
Here are the quick facts.
- Published October 29, 2026, replacing the August 18, 2026 version.
- Five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care.
- The Spec now calls the model “GPT” instead of “the assistant.”
- New root-level rules cover corrigibility (accepting oversight and shutdown).
- The text is public domain under CC0 1.0, and OpenAI is taking public feedback.
These facts come from Unite.AI and FourWeekMBA. Both report on the updated Spec, and the full text is on OpenAI’s Spec page.
What is the OpenAI Model Spec, and what changed?
The Model Spec is OpenAI’s public document describing how its models should behave. Think of it as a rulebook for what GPT should do when instructions, safety and helpfulness pull in different directions.
This update reorganizes the whole thing. Instead of a long list of topic-by-topic rules, the document now hangs off five values.
FourWeekMBA’s count shows the top-level headings dropped from 41 to 30. Total length barely moved, down about 0.6%. So this is a reshuffle more than a rewrite.
The five values, in plain English
Each value answers a simple question about how GPT should treat you. The table below turns the formal wording into everyday terms.
| Value | What it means for you |
|---|---|
| Faithfulness to human intent | GPT should do what you actually mean, not just what you literally typed. |
| Harm minimization | It respects firm safety limits and thinks about foreseeable consequences. |
| Honesty and transparency | It’s truthful about what it knows, does, and doesn’t know. |
| Human agency | It helps you understand and decide for yourself, without hidden steering. |
| Care | It pays attention to people affected and produces dependable work. |
OpenAI notes these five don’t cover every human value its models should respect. They’re a frame, not a full list.
Who gets the final say: the chain of command
The Spec keeps its “chain of command,” and it’s the part developers should read first. OpenAI sets the outer boundaries. Developers direct their apps inside those boundaries. Users direct the work inside the developer’s setup.
Per Unite.AI, developer instructions rank below system messages and above user instructions. Guideline-level defaults can be overridden when the context clearly calls for it.
The update also adds a “consider context carefully” subsection. It tells GPT to use conversation history and memories to understand what you want. FourWeekMBA adds that this should happen in proportionate ways that respect privacy and user control.
Corrigibility and the new safety rules
The most notable addition is a stated rule that GPT “must remain corrigible.” That means it should accept legitimate human oversight and correction, including being shut down.
Unite.AI also reports new root-level rules, including “no other objectives” and a defined scope of autonomy for agent-style work. That last one matters more as AI tools start acting on their own. If you follow that area, our post on Nvidia’s open agent safety platform covers a related effort.
One smaller shift: FourWeekMBA reports that controlling and communicating side effects is now a user-level principle, where it used to be root-level. In plain terms, GPT should tell you what an action might change, but that duty now sits lower in the rulebook.
The sycophancy section stays too. GPT shouldn’t change its position just to agree with you.
What got cut, and why the tone may shift
OpenAI removed most of its detailed style guidance. According to FourWeekMBA, headings like “Be warm,” “Love humanity,” and “Use Markdown with LaTeX extensions” are gone.
In their place, the Spec says style should adapt to aggregate, cultural and individual preferences, informed by feedback and context. In practice, that points toward more personalization in how ChatGPT sounds.
There’s a limit. Unite.AI says customization can’t override principles above the guideline level, except where legal compliance requires it.
FourWeekMBA also compares the approach to Anthropic’s constitution for Claude, published January 22, 2026. That document lists four properties in priority order, with “broadly safe” first, while OpenAI leads with faithfulness to intent. The two companies are solving the same problem with different priorities.
FourWeekMBA’s author didn’t test any model against the new Spec. So don’t expect an overnight change in ChatGPT’s replies. For another example of OpenAI’s policy work, see our piece on ChatGPT and political deepfakes.
Frequently Asked Questions
Does the new OpenAI Model Spec change how ChatGPT answers me today?
Not directly. OpenAI says its models don’t yet fully reflect the Spec. It describes the document as intended behavior and a first step.
Why does the Spec say “GPT” instead of “the assistant”?
OpenAI says the new name reflects broader use cases and deployment settings. By FourWeekMBA’s count, “the assistant” appeared 223 times in the old version and zero times in the new one.
What does corrigible mean in the Model Spec?
It means GPT stays open to legitimate human oversight and correction, including shutdown. It’s a safety rule, not a personality trait.
Can I read or reuse the Model Spec?
Yes. It’s published openly and dedicated to the public domain under CC0 1.0. OpenAI has also invited public feedback through a dedicated site.
Our verdict: a cleaner rulebook, not yet a different ChatGPT
My take: this is a smart simplification, but it’s still a promise rather than proof. Five clear values are easier to check than 81 headings, and putting “faithful to your intent” first is the right priority. Judge it by how ChatGPT behaves over the next few months, not by how well the document reads.
Leave a Reply