OpenAI has launched the ChatGPT agent, a new feature that allows ChatGPT to act independently using its own virtual computer. The agent can navigate websites, run code, analyse data, and complete tasks such as planning meetings, building slideshows, and updating spreadsheets.
The feature is now rolling out to Pro, Plus, and Team users, with access for Enterprise and Education users expected in the coming weeks.
The agent integrates previously separate features like Operator and Deep Research, combining their capabilities into a single system. Operator allowed web interaction through clicks and inputs, while deep research focused on synthesis and summarisation.
The new system allows fluid transition between reasoning and action in a single conversation.
“You can use it to effortlessly plan and book travel itineraries, design and book entire dinner parties, or find specialists and schedule appointments,” OpenAI said in a statement. “ChatGPT requests permission before taking actions of consequence, and you can easily interrupt, take over the browser, or stop tasks at any point.”
Users can activate agent mode via the tools dropdown in ChatGPT’s composer window. The agent uses a suite of tools, including a visual browser, a text-based browser, terminal access, and API integration. It can also work with connectors like Gmail and GitHub, provided users log in via a secure takeover mode.
All tasks are carried out on a virtual machine that preserves state across tool switches. This allows ChatGPT to browse the web, download files, run commands, and review outputs, all within a single session. Users can interrupt or redirect tasks at any time without losing progress.
ChatGPT agent is currently limited to 400 messages per month for Pro users and 40 for Plus and Team users. Additional usage is available through credit-based options. Support for the European Economic Area and Switzerland is in progress.
The standalone Operator research preview will be phased out in the coming weeks. Users who prefer longer-form, slower responses can still access deep research mode via the dropdown menu.
While slideshow generation is available, OpenAI noted that formatting may be inconsistent, and export issues remain. Improvements to this capability are under development.
The system showed strong performance across benchmarks. On Humanity’s Last Exam, it scored a new state-of-the-art pass@1 rate of 41.6%, increasing to 44.4% when using parallel attempts. On DSBench, which tests data science workflows, it reached 89.9% on analysis tasks and 85.5% on modelling, significantly higher than human baselines.
In investment banking modelling tasks, the agent achieved a 71.3% mean accuracy, outperforming OpenAI’s o3 model and the earlier deep research tool. It also scored 68.9% on BrowseComp and 65.4% on WebArena, both benchmarks measuring real-world web navigation and task completion.
However, OpenAI acknowledged new risks with this capability. “This is the first time users can ask ChatGPT to take actions on the live web,” the company said. “We’ve placed a particular emphasis on safeguarding ChatGPT agent against adversarial manipulation through prompt injection.”
To counter these risks, ChatGPT requires explicit confirmation before high-impact actions like purchases, restricts actions such as bank transfers, and offers settings to delete browsing data and log out of sessions. Sensitive inputs entered during takeover sessions are not collected or stored.
The new system is classified under OpenAI’s “High Biological and Chemical” capability tier, triggering additional safeguards. The company has worked with external biosecurity experts and introduced monitoring tools, dual-use refusal training, and threat modelling to prevent misuse.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.