OpenAI GPT-5 System Card

This document is the GPT-5 System Card published by OpenAI, a technical safety and governance disclosure covering the capabilities, evaluation methodology, risk classifications, and mitigation measures applied to the GPT-5 model family, including gpt-5-main, gpt-5-thinking, and gpt-5-thinking-nano. The card states that GPT-5 was evaluated against OpenAI's Preparedness Framework and internal safety policies, with the document disclosing that the model received a 'medium' rating across tracked risk categories including CBRN (chemical, biological, radiological, nuclear), cybersecurity, persuasion, and model autonomy. The card asserts that GPT-5 incorporates a unified model routing system that dynamically selects between sub-models based on task complexity, and discloses that the model underwent red-teaming, automated evaluations, and external third-party safety testing prior to deployment; the document does not assert user-facing contractual rights or financial terms, but does establish that operator and user access to model capabilities may be tiered and constrained through system prompt-level controls. The document engages frameworks including the EU AI Act, which classifies certain AI systems as high-risk and requires conformity assessments and transparency obligations; the FTC's authority over unfair or deceptive AI practices; and emerging voluntary commitments made under the White House AI Safety framework, though the card does not assert full compliance with any specific regulatory standard. Material compliance considerations include whether the disclosed evaluation methodology satisfies the EU AI Act's requirements for high-risk system documentation, how the 'medium' Preparedness rating translates to operator-level disclosure obligations, and how capability elicitation findings related to CBRN and cyberoffense interact with export control and dual-use research frameworks.