GPT-5.4 Thinking System Card

OpenAI has introduced the GPT-5.4 Thinking system card, detailing the safety framework for the latest reasoning model in the GPT-5 series. This release is significant as it marks the first time a general-purpose model from OpenAI has implemented dedicated mitigations to address high capabilities in cybersecurity.

Cybersecurity Safety Mitigations

GPT-5.4 Thinking implements specific safety mitigations designed to manage high capability in cybersecurity. This safety approach is an evolution of the techniques previously deployed for GPT-5.3 Codex within both the API and ChatGPT, now extended to a general-purpose reasoning model.

Model Positioning and Baselines

GPT-5.4 Thinking (internally referred to as gpt-5.4-thinking) serves as the current iteration of reasoning capabilities within the GPT-5 family. Because there is no GPT-5.3 Thinking variant, OpenAI utilizes GPT-5.2 Thinking as the primary baseline for comparing performance and safety improvements.

Safety Framework Consistency

Aside from the new cybersecurity-specific protections, the comprehensive safety mitigation approach for GPT-5.4 Thinking remains consistent with the methodologies used for previous models in the GPT-5 series.