OpenAI rewrites its Model Spec around five core values

OpenAI's updated Model Spec, the public document describing how its models should reason and behave, now rests on five core values rather than a long list of rules. The values highlighted in the release include sound judgment, preservation of human agency and faithfulness to what users actually intend.

The shift from rules toward values reflects a practical problem. As models act more autonomously across long-horizon agent tasks, enumerated rules cannot anticipate every situation. A value-based spec aims to give the model principles it can generalize from. 'Faithfulness to intent' speaks directly to agent overreach: an agent told to finish a statistics project should not bypass a paywall to get there.

The update comes the same day OpenAI detailed its misalignment monitoring, which pairs stated values with enforcement and detection. Anthropic has long published a constitution-style document, so OpenAI is converging on a similar framing. The spec also arrives amid consumer-harm scrutiny, including an NYT report on ChatGPT advising a patient to keep cutting her epilepsy medication.

Skeptics will note that a spec is a statement of intent, not a guarantee: values are only as good as training and evaluation. Watch for how OpenAI measures adherence and whether third parties can test against the spec.