OpenAI has launched GPT-5.4, its most capable frontier model yet, available in ChatGPT, the API, and Codex. The new model combines advanced reasoning, coding, and agentic workflows into a single system designed for professional knowledge work. GPT-5.4 Thinking in ChatGPT now provides upfront planning, allowing users to adjust responses mid-stream for better alignment with their needs.

On professional tasks, GPT-5.4 achieves state-of-the-art results, matching or exceeding industry professionals in 83.0% of comparisons on GDPval, which tests 44 occupations across nine major U.S. industries. The model shows particular strength in spreadsheet modeling, achieving 87.3% accuracy on investment banking analyst tasks, and creates presentations rated 68.0% higher than GPT-5.2 for aesthetics and visual variety. GPT-5.4 also reduces factual errors by 33% for individual claims and 18% for full responses.

GPT-5.4 introduces native computer-use capabilities, achieving a 75.0% success rate on OSWorld-Verified desktop navigation—surpassing human performance. The model supports up to 1M tokens of context and uses significantly fewer tokens than GPT-5.2, translating to faster speeds and reduced costs. With improved visual perception scoring 81.2% on MMMU-Pro and enhanced document parsing, GPT-5.4 delivers more reliable agents and higher-quality outputs across professional workflows.