OpenAI rewrites its Model Spec around five values and drops most style rules

OpenAI published a revised Model Spec on 29 October 2026, the public document that describes how it wants its models to behave. In its announcement, OpenAI says the updated Spec is now organized around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. Faithfulness to human intent is framed as the foundation, meaning a model should work out what OpenAI, developers and users actually want rather than follow instructions word for word.

Key facts

• Primary: OpenAI, 29 October 2026, plus the Model Spec changelog on GitHub
• Independent: Reuters via U.S. News and Aivio News, 29 October 2026
• New subsection "Consider context carefully" covers how conversations, memory and personalization should shape a model's reading of what a person wants, while respecting privacy and user control
• New defaults on supporting human agency and following social norms, plus explicit guidance on supporting human oversight and correction
• "Control and communicate side effects" becomes a user-level principle with clearer rules on authorization and trade-offs
• Most of the detailed style guidance is retired, though defaults such as approaching questions objectively stay
• The Spec now calls the model "GPT" instead of "the assistant"

How to read it

This is a statement of intent, not a product change. OpenAI says plainly that its models do not yet fully reflect the Spec, and that it is testing new training methods meant to teach the values themselves rather than a list of rules, with results promised in later updates. Reuters notes the approach echoes Anthropic's written constitution for Claude. The most concrete shift is the move away from style rules: tone will now be shaped mainly by feedback and context, so any change in how ChatGPT sounds will show up in behavior before it shows up in text anyone can read. The objectivity default is worth watching, since it is the line meant to stop a model from telling people what it thinks they want to hear. OpenAI is also asking for public feedback on the values and trade-offs.

Independence

Frontier Flux is not affiliated with OpenAI, Reuters, U.S. News or Aivio News. Descriptions of the Spec come from OpenAI's own documents and have not been tested against model behavior by Frontier Flux.