OpenAI dropped its biggest update yet today: ChatGPT Agent, an AI assistant that doesn't just chat — it actually gets stuff done. The new system can plan your dinner party, book restaurant reservations, analyze your upcoming meetings, and even create slide decks, all by controlling its own virtual computer while you sit back and watch.
Key Points
- ChatGPT Agent can browse sites, run code, and create editable files while asking for your approval on big moves
- It scores big on benchmarks, like 41.6% on Humanity's Last Exam, doubling previous models
- Safety features include permission prompts and "Watch Mode," but OpenAI warns it's experimental
- Available to Pro, Plus, and Team users; Enterprise gets it this summer
OpenAI’s new ChatGPT Agent is live—and it’s a major leap from the chatbot you’ve been chatting with. This isn’t just about answering questions anymore. It’s about taking action.
The agent can now perform full workflows across the web: it clicks buttons, fills out forms, scrapes data, runs code, and even builds slide decks. You talk to it in natural language, it figures out what to do, and gets to work using a virtual computer that mimics how a person uses a browser and desktop. OpenAI says you can ask it to brief you on calendar events based on your inbox, plan multi-stop vacations, or create financial models in spreadsheets.
It combines the muscle of OpenAI’s earlier “Operator” and “Deep Research” tools, but adds a lot more. It can switch between a visual browser, text browser, terminal, and APIs—and adapt its tools based on the task. Plus, with connectors, it can tap into Gmail, Google Drive, GitHub, and more.
This launch pushes OpenAI deeper into the AI agent race, alongside Google, Meta, and Anthropic. But where others are still in demo land, OpenAI is putting a real agent into people’s hands. It’s available starting today for paid ChatGPT users. You just pick “agent mode” and describe what you want done.
Safety is big concern that OpenAI is taking seriously—obviously an AI that can actually do things on your behalf poses fundamentally different risks than one that just talks. The company says that they trained the agent to explicitly ask for your permission before taking actions with real-world consequences, like making a purchase, and there's something called Watch Mode where if the user navigates away from the tab when the agent is operating on sensitive websites like financial portals, the agent automatically pauses execution.
The company has also activated what it calls its strongest safety measures yet. With the model's increased capabilities, we've made the decision to treat ChatGPT agent as High Biological and Chemical capabilities under our Preparedness Framework, activating the associated safeguards — a level typically reserved for AI that could potentially be misused for harmful purposes.
And while the tech is still a little slow—some tasks take up to 30 minutes—OpenAI’s team says that’s fine. You’re not meant to sit and watch. Think of it as background automation that gets smarter the more you use it.
Right now, the agent is limited to Pro, Plus, and Team users in most regions. Enterprise and Education are next. Europe still has to wait. But if you’ve been waiting for your digital assistant to finally do things instead of just talk about them, it just arrived.