OpenAI has released GPT-5.4 Thinking, the latest reasoning model in the GPT-5 series. This model implements cybersecurity mitigations for the first time in a general-purpose model, building on approaches from GPT-5.3 Codex. The safety framework follows similar patterns to previous models in the series.

The model, referred to as gpt-5.4-thinking, represents a significant step in AI safety for high-capability systems. OpenAI's approach to cyber safety evolved from techniques used in GPT-5.3 Codex and integrated into ChatGPT and the API. This development addresses growing concerns about AI systems' potential misuse in cybersecurity contexts.

GPT-5.4 Thinking serves as the primary baseline model, with GPT-5.2 Thinking as the main comparison point since no GPT-5.3 Thinking exists. The comprehensive safety mitigation approach demonstrates OpenAI's commitment to addressing capability risks while maintaining model performance. This release marks an important milestone in responsible AI development.