OpenAI GPT-5 System Card

This document describes how OpenAI has built, restricted, and structured the GPT-5 model family. A real-time router—not the user—decides which model handles each query based on conversation type, complexity, and prompt language, and once usage limits are hit, queries are automatically routed to less capable mini versions. OpenAI has placed its most powerful thinking model under heightened restrictions in biological and chemical topics as a precaution, even without confirmed evidence of harm.

This system card establishes OpenAI's safety classifications, training architecture, content moderation baseline, and deployment structure for the GPT-5 model family. OpenAI has formally classified gpt-5-thinking as High capability in the Biological and Chemical domain under its Preparedness Framework and has activated the associated safeguard set, with that classification resting on a precautionary policy choice rather than confirmed harm evidence. Safe-completions safety training is applied universally across all GPT-5 models as the baseline for preventing disallowed content. The system card's primary evaluations cover gpt-5-thinking and gpt-5-main, with other model evaluations placed in the appendix, and OpenAI discloses a near-term plan to consolidate the current multi-model architecture into a single model.

Users of GPT-5 products do not directly select which model processes their queries; OpenAI's router makes that choice based on factors including what users type in their prompts, and their switching and preference behavior feeds back into ongoing router training. When usage limits are reached, OpenAI automatically routes subsequent queries to a mini version of the relevant model with no opt-out described in this document. The gpt-5-thinking model operates under active safeguards in biological and chemical topics, meaning certain queries in those domains will be restricted regardless of user intent. The gpt-5-thinking-pro mode available in ChatGPT is a setting applied to gpt-5-thinking using parallel test time compute, not a distinct model.