OpenAI just dropped a major update to ChatGPT—and it’s a game-changer.
The new ChatGPT Agent goes beyond conversation. It can now think through tasks and take real-world actions on your behalf using a virtual computer. From booking ingredients to analyzing competitors, it doesn’t just give you advice—it gets the job done.
Here’s what’s new, what it means, and why this shift is a big deal for the future of AI tools.
What Can ChatGPT Agent Actually Do?
With the new agent mode, ChatGPT can now:
- Use the web: It browses websites, clicks links, logs in (securely), and filters information.
- Run code and work with files.
- Create editable outputs like spreadsheets and slideshows.
- Work with apps like Gmail or GitHub via built-in connectors.
- Manage workflows from start to finish—no need to jump between tools.
You can now ask things like:
- “Summarize my upcoming client meetings based on current news.”
- “Create a competitive analysis and make a presentation.”
- “Plan and shop for a dinner party.”
It will figure out how to do it—asking for your confirmation where needed—and return polished results.
What Powers This New Capability?
This upgrade merges the strengths of two earlier tools:
- Operator: An AI that could click and type on websites.
- Deep Research: A model that synthesized and summarized complex information.
Before, these tools were separate. Operator could interact with the web but wasn’t great at in-depth analysis. Deep Research could think deeply but couldn’t navigate sites. Now, ChatGPT Agent brings them together, plus more—with browsing, coding, APIs, and even terminal commands rolled into one unified system.
How It Works Behind the Scenes
ChatGPT Agent uses a virtual computer, allowing it to:
- Keep context across tasks.
- Choose between a visual browser (great for clicking around) or a text-based one (better for reading and reasoning).
- Use APIs, run code in a terminal, and interact with files.
- Seamlessly shift between thinking and doing—making decisions along the way.
It can even pause and resume tasks later, remember what it was doing, and adjust to new instructions mid-way.
A Smarter, More Collaborative AI
What makes this different from traditional AI assistants? It’s interactive and flexible.
- You’re in control: The agent always asks before taking sensitive actions, like making purchases or logging in.
- You can jump in anytime: Take over the browser, change the task, or ask for a progress update.
- It asks back: If it’s unsure, ChatGPT will request clarification to better align with your goals.
Plus, if you have the ChatGPT mobile app, it’ll notify you when tasks are complete.
Real-World Use Cases: Personal & Professional
Here’s how users are already leveraging ChatGPT Agent:
At Work
- Turn dashboards into clean, editable presentations.
- Rearrange calendars or plan team offsites.
- Automate repetitive financial modeling or spreadsheet updates.
At Home
- Plan vacations and book everything.
- Design dinner parties, shop for ingredients, and send invites.
- Schedule appointments or summarize your inbox.
How Well Does It Perform?
OpenAI tested the new agent across several expert-level benchmarks. Highlights include:
- Humanity’s Last Exam: ChatGPT Agent achieved a new high score, performing better than past models on expert questions.
- FrontierMath: On complex math problems, it significantly outperformed previous generations.
- DSBench & SpreadsheetBench: It beat human experts and AI tools alike in data science and spreadsheet tasks.
- Real-world work simulations: From financial modeling to researching new energy sites, the agent performed as well or better than skilled professionals in about half the tests.
It also set new records on tasks like web browsing, document generation, and interacting with live data.
What About Privacy and Safety?
Because the Agent can take real-world actions, OpenAI has added a robust safety system:
- Explicit permission before anything impactful happens.
- Secure login and session controls: You can delete browsing data or log out with one click.
- Privacy safeguards: Your inputs during browser sessions (like passwords) are never stored or seen by the model.
- Limitations on risky tasks: No bank transfers or other high-risk actions allowed.
OpenAI has also strengthened defenses against prompt injection attacks—when a malicious site tries to trick the AI into doing something it shouldn’t.
Special Focus: Biosecurity & Ethical Use
Given its new capabilities, OpenAI has categorized this model as high-risk for biological and chemical misuse, even though there’s no evidence it can cause harm yet. They’re being cautious.
- New safeguards for life sciences tasks.
- In-house and third-party evaluations.
- Ongoing collaboration with global biosecurity and safety experts.
This model has OpenAI’s most comprehensive safety framework to date.
Who Can Use It?
Starting now:
- Pro users get access immediately (with 400 messages/month).
- Plus and Team users will be onboarded over the next few days.
- Enterprise and Education users will follow in the coming weeks.
Note: This feature isn’t yet available in the European Economic Area or Switzerland.
The older Operator preview will shut down soon, as its capabilities are now fully rolled into the Agent.
What’s Still in Progress?
While the Agent is powerful, it’s not perfect:
- Slideshow creation is still in beta. The slides may look basic, and formatting might not always match exports.
- Spreadsheet editing is great, but slide editing from existing templates isn’t fully supported yet.
- Expect minor bugs and missteps as the system continues to improve.
OpenAI is actively developing the next generation, with smoother workflows, better formatting, and even more advanced reasoning in the works.
This update transforms ChatGPT from a smart assistant into a real agent—able to plan, reason, act, and deliver results across your digital life.
It’s the beginning of a new era of AI: one that can work with you and for you, with built-in safety and flexibility.
Read the full announcement by OpenAI here: Introducing ChatGPT Agent
