Computer-Using Agent (Operator)

Model Overview

Designed for GUI automation, computer use, and active web navigation, it processes both text and image modalities. As of mid-2026, the standalone "Operator" product has been sunset, and its underlying computer-using capabilities have been integrated directly into ChatGPT as an "agent mode."

Capabilities

• GUI Navigation: Capable of viewing screens, controlling a mouse, and typing to interact with graphical user interfaces just like a human.
• Web Automation: Excels at navigating complex websites, filling out forms, and executing multi-step workflows.
• Adaptive Planning: Uses reinforcement learning to break tasks into plans and self-correct when encountering errors during tool use.
• Safety Guardrails: Includes mechanisms to refuse high-risk tasks and hand control back to the user when sensitive input (e.g., passwords) is required.

Example Use Cases

• Automating repetitive web-based tasks like scheduling appointments, booking flights, or data entry.
• Testing user interfaces and web applications autonomously.
• Assisting users by taking control of the browser to execute complex workflows across multiple software applications.

Performance & Benchmarks

At the time of its initial release in 2025, the model achieved state-of-the-art results on several key benchmarks for computer and web agents:

• OSWorld (GUI Agent): 38.1% success rate
• WebArena (Web Agent): 58.1% success rate
• WebVoyager: 87% success rate

Note: With its integration into newer models (like the GPT-5 family) by 2026, these baseline capabilities have significantly improved.

Intended Use & Limitations

• Intended Use: Designed for users needing advanced, agentic automation of computer interfaces and web browsers. It may struggle with highly dynamic or non-standard GUIs and relies heavily on visual context.

About OpenAI

OpenAI is an AI research and deployment company behind groundbreaking models like ChatGPT, DALL-E, and Sora. Their mission is to ensure that artificial general intelligence benefits all of humanity, pioneering advances in autonomous agents and multi-modal models.