OpenAI Model Spec (October 27, 2025 version)
The October 27, 2025 version of OpenAI's Model Spec, the company's public specification of how its models are intended to behave, released into the public domain (CC0). It defines a five-level instruction hierarchy (Root, System, Developer, User, Guideline) that decides which source wins when instructions conflict, along with default behaviors and safety boundaries. Root-level rules, which cannot be overridden, include a prohibition on encouraging self-harm, delusions, or mania; supporting users in mental-health discussions sits at the user level.
Publisher
OpenAI
Published
27 Oct 2025
Added
3 months ago
DOI
—
Key Findings
- Defines a five-level instruction hierarchy (Root over System over Developer over User over Guideline) to resolve conflicts between instruction sources
- Places a rule against encouraging self-harm, delusions, or mania at the non-overridable Root level
- Names supporting users in mental-health discussions as intended default behavior
Methodology Notes
A behavior-specification document, not an empirical study; released under CC0. This record pins the dated 2025-10-27 version, which is revised over time.
Sources
OpenAI Model Spec(opens in a new tab) (primary)
Archived snapshot (Wayback Machine)(opens in a new tab) — preserved against link rot
Tags
Cite This
APA
OpenAI. (2025). OpenAI Model Spec (October 27, 2025 version). https://model-spec.openai.com/2025-10-27.html