GPT-5 System Card

GPT-5 is an integrated system combining gpt-5-main for fast responses and gpt-5-thinking for deep reasoning. A real-time router analyzes conversation complexity, tool needs, and user intent to automatically select the optimal model, and continuously learns from user feedback.

The model lineup is segmented by use case as follows:

• gpt-5-main and gpt-5-main-mini: high-speed/high-throughput models
• gpt-5-thinking, gpt-5-thinking-mini, gpt-5-thinking-nano: reasoning-specialized models
• gpt-5-thinking-pro: a version that leverages parallel test-time compute in ChatGPT

This system reduces hallucination compared to previous models and greatly improves instruction following and resistance to sycophancy. In particular, performance has significantly improved in writing, coding, and health, and all models have adopted a new safety training method called safe-completions.

On the safety front, gpt-5-thinking has been classified as High capability in biology and chemistry, activating related safeguards. This is a precautionary measure against the possibility that the model could help cause high-level biological harm.