OpenAI Releases GPT-5.4 AI Models With Agentic Computer-Use Capabilities

OpenAI released the GPT-5.4 artificial intelligence (AI) model on Thursday. The third major update to the GPT-5 foundational model arrives with improvements across parameters and new agentic capabilities, including computer-use. Unlike GPT-5.3-Codex, which was only released as a coding model for the Codex platform, the latest large language model (LLM) is being released for both ChatGPT and the application programming interface (API). Additionally, the San Francisco-based AI giant also released a more advanced version of the model dubbed GPT-5.4 Pro.

GPT-5.4 AI Models Released

In a post, the AI giant announced and detailed the release of the new AI models. GPT-5.4 is available in both Codex and API, as well as in ChatGPT as GPT-5.4 Thinking. At launch, the model will only be available to the Plus, Team, and Pro subscribers, replacing GPT-5.2 Thinking. On the other hand, the GPT-5.4 Pro is currently available only to Pro and Enterprise subscribers.

OpenAI says that the GPT-5.4 model brings improvements in reasoning, coding, and agentic workflows. It incorporates the coding capabilities of 5.3-Codex, as well as improves general-purpose behaviour when working with tools and carrying out tasks involving spreadsheets, presentations, and documents.

However, the most interesting new capability is computer-use. It refers to an LLM's ability to deploy AI agents that can access a virtual machine and carry out complex tasks involving web browsing, placing orders, and more. OpenAI claims GPT-5.4 can write code to operate computers, issue mouse and keyboard commands, and has a steerable behaviour.

Based on internal evaluations, the AI giant claimed that the model scored a 75 percent success rate on the OSWorld-Verified benchmark test, which measures a model's ability to navigate a desktop environment through screenshots and keyboard and mouse actions.

OpenAI says the model's improved computer-use skills are powered by its enhanced visual perception capabilities. On the MMMU-Pro benchmark, which measures an AI model's visual reasoning, GPT-5.4 scored an 81.2 percent success rate.

Coming to safety, the company said GPT-5.4 is classified as High cyber capability, and is being deployed with an expanded cyber safety stack that includes monitoring systems, trusted access controls, and asynchronous blocking. OpenAI has also introduced chain-of-thought (CoT) controllability as a new open-source evaluation. It measures whether models can deliberately obfuscate their reasoning to avoid being monitored.