OpenAI updates Model Spec to guide more capable AI

OpenAI said its updated Model Spec is intended to guide more capable AI toward sound judgment, human agency and faithfulness to human intent, according to the company’s announcement published on 2026-10-29.

Evidence and context

The source material says the revised specification lays out five core values, but it does not enumerate them in the provided excerpt. OpenAI’s Model Spec is its public, evolving description of intended behavior for models used in its products and API, and the document is meant to govern how models should respond across areas such as usefulness, safety, honesty, uncertainty, privacy and harmful requests. The background material also notes that the spec is not itself a model release and does not guarantee deployed systems already comply with every rule.

Related official OpenAI material shows the specification has been updated repeatedly. OpenAI published model release-note entries dated 2025-09-12 and 2025-10-27, and an official announcement dated 2025-12-18 added teen protections covering self-harm, sexualized or violent immersive roleplay, dangerous activities, substance misuse and concealing harm. The background provided also says the document’s instruction hierarchy gives root, system, developer, user and guideline instructions different levels of priority when they conflict.

The significance of the latest update is that it frames model behavior around broader judgment and alignment principles as capabilities increase. The provided evidence does not state whether the 2026-10-29 update changed production behavior, nor does it establish which version is currently the live reference beyond noting that the official documentation landing page is the reference for the current version.