OpenAI Launches ChatGPT Agent With Virtual Computer and Web Access [Video]

OpenAI has unveiled ChatGPT agent, a new capability that allows the popular AI to go beyond conversation and take action on a user's behalf. Activated via a new "agent mode," which is rolling out to paid subscribers, the feature gives ChatGPT its own virtual computer to handle complex, multi-step tasks from start to finish.

The new system combines the web-browsing abilities of its previous "Operator" tool with the synthesis and analysis skills of "deep research," allowing the AI to intelligently navigate websites, run code, and deliver editable documents. According to OpenAI, users can now make requests such as "plan and buy ingredients to make Japanese breakfast for four" or "analyze three competitors and create a slide deck," though the company notes the slideshow capability is still in beta and may lack polish.




At its core, ChatGPT agent uses a suite of tools including a visual browser, a text-based browser for simpler queries, a terminal, and API access through ChatGPT connectors. This allows the model to choose the most efficient path for a given task, whether it's pulling information from a connected Gmail account or running a command in its virtual computer. OpenAI emphasizes that users remain in control, with the ability to grant permissions, take over the browser, or stop a task at any time.

This software launch comes as OpenAI pushes into hardware on a separate front. The company recently acquired io, the AI hardware startup from former Apple design chief Jony Ive. Reports suggest they are working on a AI device with no screen that could enter mass production in 2027.

This latest release further intensifies the AI race. Apple is currently working to deliver on its own Apple Intelligence platform, though its most advanced, personalized Siri features have been delayed until 2026. In a recent interview, Apple executives explained that this was a necessary shift to a more robust underlying architecture. While Apple has integrated ChatGPT into iOS for some tasks, this new agentic capability from OpenAI operates independently.

OpenAI acknowledged that these new capabilities introduce new risks, particularly around prompt injection and the handling of sensitive data. Reflecting the agent's advanced capabilities, the company is treating it as having "High Biological and Chemical capabilities" under its internal safety framework, activating its most comprehensive safeguards to date. This includes requiring explicit user confirmation for consequential actions and training the agent to refuse high-risk tasks like bank transfers.


OpenAI also notes that the workflow is designed to be collaborative. Users can interrupt tasks to provide new instructions, and the agent will adapt without losing progress. It can also be paused, asked for a summary of its work, or stopped entirely, with a notification sent to the user's phone via the ChatGPT app when a task is complete.

The company backed the launch with impressive performance benchmarks. The model powering ChatGPT agent set a new state-of-the-art score on Humanity's Last Exam, an evaluation measuring AI performance on expert-level questions. It also significantly outperformed human baselines on the DSBench benchmark for data science tasks and showed major gains in an internal test of investment banking modeling tasks.

ChatGPT agent is rolling out starting today for Pro, Plus, and Team users, who can enable it through the tools dropdown in the message composer. Pro users will receive 400 messages per month, while other paid subscribers get 40, with options for additional usage through credit-based plans.

Please download the iClarified app or follow iClarified on Twitter, Facebook, YouTube, and RSS for more AI news.