Illustration, AI-generated from the story text. Tap a panel to enlarge.
OpenAI rewrites its Model Spec around five values and tells agents to ask before spending
OpenAI published a new version of its , the rulebook for how its models should behave, dated October 29th and "organized around five values": be faithful to human intent, minimize harm, be honest and transparent, facilitate human agency, and act with care. Remembered preferences "are defaults and should be respected when they make sense in the current context," and cannot override an explicit instruction.
The rules are the practical change. A model should propose terms and get approval before buying something; inside an authorized budget it may proceed, but a cheaper nonrefundable fare needs separate permission. A model should raise an unsolicited concern once, then stop repeating it. OpenAI says its models do not fully meet the new spec yet and gave no timetable. SourcesAC
Sources
- AOpenAI Model Spec, October 29th 2026model-spec.openai.com
- COpenAI's Model Spec and human intent (The Neuron)theneuron.ai
This story ran in the Morning Brief of October 30th 2026.