AIImpish
On October 29, 2026, OpenAI published a major update to its Model Spec, the public document that describes how its models are supposed to behave. Instead of a long list of rules and style tips, the Spec is now organized around five core values, and it calls the model "GPT" rather than "the assistant."
Coffee summary
- FACT: The Spec now rests on five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care.
- FACT: OpenAI added new guidance on reading context (conversations, memories), supporting human agency, following social norms, and handling side effects.
- FACT: Most of the detailed style guidance is retired; style should adapt to each person's preferences.
- FACT: OpenAI says its models "do not yet fully reflect" the new Spec.
- OPINION: This is a document about direction, not a switch that changes ChatGPT overnight.
What the Model Spec is
Think of the Model Spec as the rulebook OpenAI trains toward. It sets firm boundaries for safety and then lists defaults that developers and users can change. When people ask "why did ChatGPT answer like that?", the Spec is often the closest thing to an official answer.
The five values, in plain words
| Value | What it means in practice |
|---|---|
| Faithfulness to human intent | Understand what you actually want, not just the literal words you typed |
| Minimizing harm | Keep firm safety lines and think about consequences of actions |
| Honesty and transparency | Say what the model knows, what it did, and what it is unsure about |
| Human agency | Help you decide and act for yourself instead of quietly steering you |
| Care | Pay attention to your situation and do work you can rely on |
OpenAI calls faithfulness to human intent "the foundation." Its example: when you compare job offers, the model may remember that you said you value predictable hours, but it should show that assumption so you can correct it.
What is new
- Consider context carefully: a new subsection on how chats, memories, and other clues should shape the model's picture of what you want.
- Supporting human agency: new defaults aimed at strengthening your own skills and choices.
- Side effects: expanded guidance on controlling and communicating what an action might change beyond the task itself. This matters more as models click, type, and run tasks for you.
- Updated chain of command: clearer rules about whose instructions win when OpenAI, a developer, and a user disagree.
What is going away
Most of the detailed style guidance is retired. OpenAI says style should follow what people prefer, by language, culture, and individual. One strong default stays: the model should approach questions objectively and should not guess that you want a particular viewpoint just because it thinks you would like to hear it. You can still ask for a perspective explicitly.
Limitations
CLAIM: OpenAI says it is exploring new training methods to teach these values, but it has not published results yet. The company is clear that the Spec describes goals, not current behavior.
What to do next
- Everyday users: if ChatGPT makes an assumption about you, correct it out loud; the new Spec says it should make those assumptions visible.
- Developers: reread the chain of command section, since it defines how your system prompts interact with user requests.
- Everyone: OpenAI invites public feedback on the Spec, so this is a rare chance to weigh in.
AIImpish Take
The headline is not a new feature but a new philosophy: understand intent, admit uncertainty, and leave people in charge. That matters most for agents that act on your behalf. Watch for whether future model releases actually behave this way.
*Sources: OpenAI, "An updated Model Spec for more capable and beneficial AI".*
