OpenAI has released GPT-5.4, a new flagship AI model that moves beyond simple chat responses and into real computer control.
The model—available in ChatGPT as “GPT-5.4 Thinking”—is also accessible through the OpenAI API and the company’s coding tool OpenAI Codex. A Windows version of Codex was recently introduced, allowing developers to run the AI model directly within coding workflows.
AI that can act, not just answer
The most notable feature of GPT-5.4 is its ability to interact with a computer through AI agent systems. Instead of simply explaining how to complete a task, the model can trigger actions such as clicking a mouse, typing commands, editing files, and analyzing screenshots.
These actions are executed by a local AI agent that receives instructions from GPT-5.4. In practice, that means the model can navigate software interfaces, open programs, and complete tasks within applications.
For example, users could ask an AI agent powered by GPT-5.4 to open financial software like Quicken, navigate menus, and perform bookkeeping tasks automatically.
New reasoning and planning abilities
Alongside computer-use capabilities, GPT-5.4 also introduces improvements in reasoning efficiency. According to OpenAI, the model can solve problems using fewer tokens, which reduces computing costs when using the API.
Another addition is the ability to generate an “upfront plan” before executing complex instructions. This allows users to review and adjust the strategy before the AI begins performing actions.
The model also includes improved spreadsheet capabilities, making it better suited for tasks involving data analysis and structured information.
Important limitations
Despite its expanded abilities, GPT-5.4 cannot directly control a user’s computer when accessed through the ChatGPT web or desktop interface. In those cases, it remains limited to the chat environment and integrated services such as cloud storage or creative tools.
Full computer control is only available through the OpenAI API or Codex environments where an AI agent system can execute commands on the local machine.
The release highlights a growing shift toward agentic AI, where language models collaborate with automated agents to perform real tasks on a computer rather than just providing instructions.