OpenAI Model Spec (model-spec.openai.com)

OpenAI published a public Model Spec that codifies the intended behavior, safety priorities, and operational rules for the models that power ChatGPT and the API. The spec frames high‑level “red-line” prohibitions (e.g., facilitating mass violence, WMDs, child sexual abuse, targeted manipulation or mass surveillance), commits to privacy and transparency, and explains trade-offs between maximizing helpfulness, minimizing harm, and preserving OpenAI’s license to operate. It’s meant to increase transparency, guide alignment work, and invite public feedback; the document is dedicated to the public domain (CC0) and will be iteratively updated. OpenAI also notes production models are still being refined toward full adherence to the spec. Technically the spec defines a chain-of-command for conflicting instructions, formalizes conversation structure and roles (system, developer, user, assistant, tool), and specifies settings like max_tokens and end_turn. It clarifies tool invocation semantics and cautions about irreversible side effects, documents hidden chain-of-thought (internal reasoning kept from users), and categorizes risks into misaligned goals, execution errors, and harmful instructions—each with mitigation strategies such as asking clarifying questions, expressing uncertainty, and refusing unsafe requests. For developers this means clearer defaults and boundaries for customization, new expectations about how to design tool-enabled agents, and improved interpretability of model behavior as OpenAI incrementally aligns deployments with these rules.