OpenAI launches ChatGPT agent – here’s everything you need to know

OpenAI has officially launched ChatGPT Agent, a major upgrade to its popular AI assistant that pushes the boundaries of what artificial intelligence can do in real-world scenarios.

Starting today, Pro, Plus, and Team users can activate the new capabilities, which allow ChatGPT to handle end-to-end tasks across the web, conduct complex research, automate workflows, and even deliver editable files, all powered by its virtual computer.

This release marks a significant evolution in AI, from passive conversation to active participation in your work and life.

What Is ChatGPT Agent?

ChatGPT Agent introduces a unified “agentic” system that combines the strengths of OpenAI’s prior tools, Operator, Deep Research, and the core ChatGPT intelligence into a single powerhouse. It can now:

• Browse websites interactively
• Run code via a built-in terminal
• Use APIs and app connectors (e.g., Gmail, Google Calendar, GitHub)
• Prompt secure logins for authenticated sessions
• Create editable spreadsheets and slides
• Schedule and automate recurring tasks

Imagine asking ChatGPT to analyze competitors and build a pitch deck, or to plan a trip, book reservations, and send confirmations, and it delivers, complete with visual narration of every step.

Seamless Transition from Conversation to Action

One of the standout features is how fluidly you can shift from chatting with ChatGPT to giving it real tasks. Whether it’s researching market trends, formatting a financial model, or planning a dinner party, the model can interpret instructions, adapt mid-task, ask for clarification, and even notify you via the mobile app once it’s done.

This makes ChatGPT Agent ideal for collaborative, iterative workflows. You’re always in control: you can pause, take over, or get a task summary at any point.

Tools Behind the Agent

The agent comes equipped with:

• Visual browser: Mimics how a human would navigate a site
• Text browser: Efficient for information extraction and reasoning
• Terminal: Executes code for data analysis or problem-solving
• API and connectors: Pulls live data from your apps like Gmail, Calendar, Notion, etc.

These tools are orchestrated through ChatGPT’s virtual computer, which allows multi-step tasks to be executed in context, efficiently and accurately.

Real-World Use Cases

From automating spreadsheet edits and dashboard reports, to booking offsites, to writing detailed investment models, ChatGPT Agent is built to handle tasks across domains:

• Business: Research, analysis, project planning, calendar management
• Finance: LBO models, amortization schedules, expense processing
• Personal life: Travel planning, event organizing, meal prepping

It’s an all-in-one AI worker that adapts to your unique workflow.

Benchmarking Performance: A New Standard for AI

ChatGPT Agent doesn’t just promise capability. It’s backed by state-of-the-art (SOTA) performance:

• Humanity’s Last Exam (HLE): 41.6 pass@1, rising to 44.4 with parallel task attempts
• FrontierMath: 27.4% accuracy on expert-level math problems
• DSBench: Outperformed human-level performance on realistic data science tasks
• SpreadsheetBench: 45.5% vs Copilot’s 20.0% accuracy
• BrowseComp: 68.9%, 17.4 points ahead of prior models
• WebArena: Best-in-class web task execution

It also outshines human output in many knowledge-intensive, economically valuable tasks across industries.

Safety, Privacy, and Control

Given its power, ChatGPT Agent introduces novel risks, and OpenAI has responded with robust safeguards:

• Explicit user confirmations for high-risk actions
• Active supervision (“Watch Mode”) for tasks like sending emails
• Refusal training for unsafe tasks (e.g., bank transfers)
• Privacy-first browsing with no data stored during browser sessions
• Easy one-click deletion of browsing data and session logouts
• Biological and chemical risk controls, treating the model as high-risk under OpenAI’s Preparedness Framework

A key focus is resisting prompt injection attacks, where malicious web elements could mislead the agent. OpenAI has built proactive detection, training, and oversight mechanisms to counter these vulnerabilities.

Availability and Access

Starting July 17, Pro users gain immediate access (400 messages/month), while Plus and Team users (40 messages/month) will onboard in the coming days. Enterprise and Education users are next in line. Flexible credits are also available for additional usage.

According to Open AI, access in the European Economic Area and Switzerland is still pending, and the Operator research preview site will be retired soon.

Subscribe to Notifications