July 18, 2025 — San Francisco — OpenAI has officially launched its most advanced and capable AI assistant yet: the ChatGPT Agent, a unified system designed to handle complex, long-running tasks autonomously using a virtual computer, suite of tools, and deep collaborative behavior.
During a live demonstration, OpenAI’s leadership, including CEO Sam Altman, showcased the full capabilities of the new Agent system, which merges the strengths of its earlier tools — Deep Research and Operator — into a single, powerful assistant. The goal: create an AI that not only thinks and reasons, but also acts, using tools like a browser, terminal, and file editors to get real-world work done on a user’s behalf.
A Unified, Tool-Using AI System
Launched initially for Pro, Plus, and Team ChatGPT users, the ChatGPT Agent combines three key interfaces inside a virtual machine:
- Text Browser: Allows for efficient and rapid reading and analysis of web content.
- GUI Browser: Provides interaction with websites just like a human would — clicking, scrolling, and filling out forms.
- Terminal: Executes code, analyzes data, generates files like slides and spreadsheets, and calls APIs to access services like Google Drive, Calendar, GitHub, and more.
“We launched Deep Research and Operator earlier this year, and the reception was incredible,” said Altman. “But users wanted something more — a single agent that could handle all kinds of tasks, seamlessly and autonomously.”
Now, that vision has materialized.
A Wedding Plan That Plans Itself
To illustrate its functionality, the OpenAI team showed a scenario where the Agent planned an entire wedding attendance: selecting outfits based on dress codes, booking hotels, and even suggesting thoughtful gifts. With just one prompt, the Agent launched its virtual computer, browsed the web for options, used visual previews, checked availability, and generated a polished itinerary.
“It was remarkable to watch the Agent just do all the work,” said a team member during the stream. “Something that usually eats up hours was reduced to minutes.”
The Agent can also be interrupted mid-task to take new instructions — such as adding a shoe shopping task — without losing context, a major upgrade from previous AI models. This multiturn interactivity mimics a real assistant, who might ask for clarifications, confirmations, or updates while working.
Deep Collaboration and Intelligent Tool Use
The new model has been trained using reinforcement learning, where it learned not just how to use tools, but when to use which tool depending on the complexity of the task. For example:
- Making restaurant reservations: Begins with deep research via the text browser, then switches to GUI for booking.
- Creating a design artifact: Searches online resources, writes code in the terminal, and compiles output in visual editors.
- Generating business reports: Connects to user data (with permission), creates spreadsheets, and delivers polished presentations.
One demo showed the Agent creating a PowerPoint presentation entirely from scratch by pulling evaluation data from internal APIs, generating charts, and decorating slides using image generation models.
Setting the Bar on Benchmarks
In benchmark tests, the ChatGPT Agent model has set new state-of-the-art results:
- 42% on Humanities Last Exam with tool use — nearly double previous models.
- 27% on FrontTMS for mathematical reasoning.
- 69% pass rate on BrowseComp (vs. 46% for previous models).
- 45% success rate on real-world spreadsheet tasks.
- Outperforms all previous models on investment banking simulations.
“This is one of the most powerful models we’ve ever trained,” said the OpenAI team. “It’s not just smart, it’s capable.”
Guardrails and Security in a New AI Era
As revolutionary as the technology is, OpenAI emphasized the risks and responsibilities that come with it. One critical area of concern is prompt injection attacks, where malicious websites try to manipulate the agent into taking unsafe actions.
To combat this, OpenAI has:
- Trained the Agent to ignore suspicious web content.
- Implemented real-time monitoring systems that can pause dangerous behavior.
- Introduced takeover mode, allowing users to take direct control of the browser for sensitive actions (e.g., entering credit card information).
“This is a new surface — with new risks,” said Chief Scientist Ilya Sutskever. “Just like the internet brought new threats, so too will AI agents. But we’ve built robust defenses, and we’ll evolve with the threat landscape.”
Available Now for Pro and Team Users
The ChatGPT Agent is now rolling out to Pro users (400 queries/month) and Team users (40 queries/month). Enterprise and Education rollouts are expected by the end of the month.
“Watching it navigate the web, make bookings, generate spreadsheets, and collaborate in real-time — this is not just the future of AI, it’s the future of getting things done,” Altman concluded.
Wrapping up
The ChatGPT Agent represents a major leap toward autonomous digital assistance. It bridges the gap between passive information retrieval and active task execution — reshaping how we interact with AI.
With safety protocols in place and continual improvement ahead, OpenAI’s Agent could become the most capable, helpful, and collaborative AI assistant yet.
Stay updated with the latest in AI and tech by following UState Pulse—your go-to source for cutting-edge insights and breaking news.