Sharing the latest Model Spec

Source: OpenAI Date: 2025-02-12 URL: https://openai.com/index/sharing-the-latest-model-spec

Summary

OpenAI’s February 2025 update to its model spec — the governing document that defines how ChatGPT and API models are expected to behave, prioritize competing goals, and handle edge cases. This revision came roughly nine months after the original model spec publication (May 2024) and addressed areas where the spec needed refinement based on deployment experience. Notable additions included teen safety protections (formalized in December 2025 as a separate update) and clearer handling of operator vs. user trust hierarchies.

Implications

Model spec as a living governance document. The February 2025 revision signals that OpenAI is treating the model spec as an iterative policy instrument rather than a one-time publication. That’s the right approach for a system deployed at scale, but it also means the behavioral contracts developers and operators rely on can shift. Tracking model spec versions is now part of responsible API dependency management.

Operator trust hierarchy clarification. One of the known gaps in the May 2024 spec was how to handle conflicts between operator system prompts and user requests. The February 2025 revision likely tightened this — important for enterprise deployments where system prompts are doing significant behavioral shaping.

Thread: alignment documentation. Sits alongside the chain-of-thought monitoring paper (March 2025), the preparedness framework update (April 2025), and the safety bug bounty (March 2026) as OpenAI’s published-alignment-posture arc. The spec is the document that downstream safety researchers evaluate most carefully.

Watch: Whether the model spec’s “broadly safe” vs. “broadly ethical” priority ordering is preserved or modified — that hierarchy is the most consequential policy decision in the document.