OpenAI launched ChatGPT agent, which completes tasks on its own virtual computer. It merges Operator's web clicking with deep research's summarising. The agent can browse visually, run code in a terminal and call interfaces directly. It returns editable slideshows and spreadsheets. Users can interrupt it, take over the browser or stop a task. It asks permission before consequential steps and refuses bank transfers. OpenAI warns that hidden instructions on web pages could hijack it. It is treating the model as high capability for biological and chemical risk. Slideshow generation is in beta and formatting can be rough. Pro, Plus and Team got access first, with Enterprise and Education later.
What changed
Operator could click through websites; deep research could only read and summarise.
What it unlocks
Asking ChatGPT to browse, log in, run code and return editable slides or spreadsheets in one task.
- 400 agent messages/month on Pro
- 40 messages/month other paid plans
- 41.6 pass@1 on Humanity's Last Exam
- 68.9% on BrowseComp
What you need to act on it
- Pro, Plus or Team plan
- agent mode selected in the tools dropdown
- connectors authorised for email or calendar access
Sources