Introducing GPT-5.4

OpenAI announced GPT‑5.4, a new frontier model released in ChatGPT as GPT‑5.4 Thinking, in the API, and in Codex. The model builds on GPT‑5.3‑Codex’s coding strengths while adding native, state‑of‑the‑art computer‑use capabilities that allow agents to operate computers and complete complex workflows across applications. GPT‑5.4 supports up to 1M tokens of context, enabling longer planning and verification cycles. In benchmark tests, GPT‑5.4 achieved a 75.0% success rate on OSWorld‑Verified desktop navigation, 67.3% on WebArena‑Verified browser use, and 92.8% on Online‑Mind2Web screenshot‑based browser tasks, surpassing GPT‑5.2 and human performance in some cases. The model also shows improved visual perception, scoring 81.2% on MMMU‑Pro without tool use and reducing document parsing error to 0.109 on OmniDocBench. In professional knowledge work, GPT‑5.4 matched or exceeded industry professionals in 83.0% of GDPval comparisons, up from 70.9% for GPT‑5.2, and achieved an 87.3% mean score on spreadsheet modeling tasks. Hallucination rates dropped, with individual claims 33% less likely to be false and full responses 18% less likely to contain errors. OpenAI also launched a ChatGPT for Excel add‑in and updated spreadsheet and presentation skills in Codex and the API. Developers can access these features through updated documentation and can configure safety and tool use policies. The release positions GPT‑5.4 as the most capable model for building agents that perform real‑world tasks across websites and software systems.