OpenAI has introduced GPT‑5.4 Thinking, the newest reasoning model in its GPT‑5 series. According to the company, the model follows the series’ comprehensive safety mitigation approach and is the first general‑purpose model to incorporate mitigations for high‑capability cybersecurity. The cybersecurity strategy builds on the latest methods implemented for GPT‑5.3 Codex, as well as in ChatGPT and the API. In the card, OpenAI refers to the model as gpt‑5.4‑thinking, noting that there is no GPT‑5.3 Thinking model; therefore, GPT‑5.2 Thinking serves as the main baseline for comparison. The release highlights OpenAI’s continued focus on safety and security as it expands the capabilities of its language models.
GPT-5.4 Thinking System Card
OpenAI has released GPT‑5.4 Thinking, the newest reasoning model in the GPT‑5 series. The model follows the series’ safety mitigation framework and is the first general‑purpose model to include mitigations for high‑capability cybersecurity. The approach builds on methods used in GPT‑5.3 Codex, ChatGPT, and the API. GPT‑5.4 Thinking is also referred to as gpt‑5.4‑thinking, with GPT‑5.2 Thinking as the baseline model.
Why it matters
The inclusion of cybersecurity mitigations in a general‑purpose model signals a shift toward more secure AI deployments, addressing growing concerns about AI misuse in security contexts.
Topic dossiers OpenAI