The OpenAI Model Spec is a formal public framework that defines how the company’s AI models should behave across the wide range of queries users submit daily. OpenAI published a detailed account of the framework’s structure, philosophy, and evolution on March 26, 2026, offering transparency into how intended model behavior is defined and implemented.
The Model Spec is not a claim that current models behave perfectly according to its standards. Rather, it serves as both a description of intended behavior and a target for future improvement. OpenAI uses it to train models, evaluate outputs, and refine behavior over time.
Why OpenAI Model Spec Exists
OpenAI stated that public clarity about model behavior matters for both fairness and safety. People need to understand how and why artificial intelligence systems treat them in specific ways, and institutions need clear expectations for how capable AI systems are intended to behave. The framework has evolved substantially since its first version in 2024, incorporating user feedback and expanding to address greater model capabilities.
The company said the Model Spec complements two other initiatives: the Preparedness Framework, which addresses risks from frontier capabilities, and AI resilience efforts, which focus on broader societal challenges as advanced AI systems are deployed.
Structure and Key Components
The Model Spec opens with a preamble that outlines three system-level goals: deploying models that empower developers and users, preventing models from causing serious harm, and maintaining OpenAI’s license to operate. The preamble is not a direct instruction to the model but provides guidance for resolving ambiguity in applying the framework’s more detailed principles.
The framework also contains public commitments that extend beyond measurable model behavior. These include a commitment that first-party deployments such as ChatGPT will never use system messages to intentionally compromise objectivity, and that model responses will be optimized for user benefit rather than revenue or non-beneficial time-on-site.
The Chain of Command
At the core of the Model Spec is the Chain of Command, a mechanism for resolving conflicts between instructions from different sources, including OpenAI, developers, and users. Each policy and instruction carries an authority level, and the model prioritizes higher-authority instructions when conflicts arise.
Hard rules form one layer of this structure. These are non-overridable boundaries that prevent models from contributing to catastrophic risks, causing direct physical harm, violating laws, or undermining the chain of command itself. A separate section called Under-18 Principles adds additional safeguards for users under 18 years of age.
Defaults form a second layer. These are overridable starting points that define the model’s behavior when users or developers have not specified a preference. Some defaults, such as tone and style, can be adjusted implicitly. Others, such as truthfulness and objectivity, require explicit instructions to override, ensuring that shifts in factual stance remain transparent.
Interpretive Aids and Ongoing Evolution
Beyond the hierarchy, the Model Spec uses decision rubrics and concrete examples to help models apply principles consistently in ambiguous situations. Decision rubrics list considerations such as minimizing irreversible actions, keeping actions proportionate to the objective, and favoring reversible approaches. Concrete examples show compliant and non-compliant responses near important decision boundaries.
OpenAI said it keeps the number of examples relatively small, focusing on the most informative cases, while broader evaluation suites cover the longer tail of edge cases. The company added that it is investing in public feedback mechanisms, including collective alignment initiatives, to keep the public involved in shaping how AI behavior evolves over time.
“The Model Spec is our attempt to make intended model behavior explicit: not just inside our training process, but in a form that users, developers, researchers, policymakers, and the broader public can actually read, inspect, and debate.”
OpenAI
The framework is described as an evolving document. OpenAI stated it will continue to modify individual elements as it learns from real-world deployment and public feedback, with the goal of making the transition to advanced AI gradual, iterative, and publicly legible. The company also noted that safety and accountability mechanisms are central to this long-term approach.