A revised guide, with an implementation limit
OpenAI’s October 29 Model Spec update organizes its behavior guidance around five values, including human intent, honesty, and human agency. It adds or expands guidance on context, oversight, and side effects, while retiring much detailed style guidance. [1]
The dated Spec describes what OpenAI is training its models toward. Both sources explicitly say production models do not yet fully reflect it. The announcement does not establish that every deployed model changed behavior today. [1][2]
Our take
For general readers, this is useful context when deciding what to delegate. Make the intended outcome and permission boundaries clear, then inspect what the product actually does. A published behavior target does not by itself verify a particular action, safeguard, or answer.
We reviewed public sources; we did not test model compliance.