OpenAI Details GPT-5.4 Thinking Safety

OpenAI has published the system card for its latest reasoning model, GPT-5.4 Thinking. This new iteration in the GPT-5 series, detailed in OpenAI News, marks a significant step in AI safety by addressing high-capability risks in cybersecurity for the first time in a general-purpose model.

The safety framework for GPT-5.4 Thinking largely mirrors that of its predecessors. However, it incorporates advanced cyber safety measures, drawing from the methodologies used for GPT-5.3 Codex, which powers features in ChatGPT and the API. This approach aims to proactively manage potential misuse in the cybersecurity domain.

The model, referred to internally as gpt-5.4-thinking, is baselined against GPT-5.2 Thinking, as there was no GPT-5.3 Thinking variant. This comparison highlights the incremental safety and capability advancements within the GPT-5 family. OpenAI's ongoing commitment to safety is evident in the structured approach to deploying these powerful models.