OpenAI has introduced GPT-5.4, the latest update in the GPT-5 series. This new version offers improvements in reasoning, coding, and agent workflows, and also adds the ability for AI agents to use computers to complete tasks.

The model is now available through ChatGPT, an API for developers, and the Codex environment.

GPT 5.4 Introduces Computer-Using AI

One of the most significant additions in GPT-5.4 is the new ability to use computers. This allows the model to run AI agents in a virtual computer environment, where it can perform actions such as:

  • Browsing the web
  • Filling out forms or placing online orders
  • Editing spreadsheets, documents, and presentations
  • Writing and executing code
  • Using mouse and keyboard commands to control software

This feature brings the model closer to acting as a digital assistant capable of completing real tasks rather than just generating text responses.

New GPT Improves Reasoning and Coding

While integrating the coding capabilities introduced in GPT-5.3-Codex, GPT-5.4 also expands the model's ability to manage broader workflows.

OpenAI states that the update enhances performance in areas such as logical reasoning, software development tasks, and multi-step workflows involving external tools. The model also includes steerable behaviors that allow users to guide how the AI handles tasks.

OpenAI Introduces GPT-5.4 Pro for Advanced Loads

In addition to the standard model, OpenAI has launched a more powerful version designed for demanding enterprise and professional loads, called GPT-5.4 Pro. Access to the models is divided as follows:

  • GPT-5.4 Thinking - Available to ChatGPT Plus, Team, and Pro users
  • GPT-5.4 Pro - Available to Pro and Enterprise users

The new model replaces the previous GPT-5.2 Thinking version.

GPT 5.4 Improves Security Systems and Benchmark Results

OpenAI has also introduced new security measures for the model. GPT-5.4 has been included in the high cyber capability class; this means it now includes stronger monitoring systems and tighter access controls to help prevent misuse.

The company has also introduced a new assessment method called "mind chain controllability," designed to detect whether models are attempting to obscure their reasoning. In internal tests, GPT-5.4 achieved the following success rates:

  • 75% success rate in desktop task navigation on the OSWorld-Verified benchmark
  • 81.2% score on visual reasoning in the MMMU-Pro benchmark

OpenAI Continues Its Focus on AI Agents

With features like computer control and enhanced reasoning, GPT-5.4 marks a step closer to agent-style AI systems capable of completing complex workflows with minimal human input.

The latest release from OpenAI highlights a broader industry shift towards AI tools that can automate real-world tasks through productivity software and digital services.