OpenAI has launched GPT-5.4, a new large language model designed for enhanced automation of work tasks. Available across ChatGPT, Codex, and its API, the model significantly reduces token usage compared to its predecessor, GPT-5.2, thereby lowering inference computing costs.
GPT-5.4 streamlines application workflows by enabling the model to automatically identify and select necessary external tools for a given task. This eliminates the need for developers to manually upload extensive tool lists, reducing API request sizes and associated inference costs.
The new model supports requests with up to 1 million tokens and offers significantly improved image processing capabilities. Developers can now upload high-resolution images exceeding 10 million pixels without compression, preserving critical details.
With upgraded vision capabilities, GPT-5.4 demonstrates superior computer vision performance. It achieved an industry-record score of 75% on the OSWorld-Verified benchmark, surpassing both GPT-5.2 and typical human tester performance.
Further enhancements include an over 8% improvement on spreadsheet analysis benchmarks and better performance in presentation preparation, online research, and science question answering. GPT-5.4 is accessible via API at $2.5 per million input tokens and $12 per million output tokens, with an advanced edition, GPT-5.4 Pro, available for complex tasks.