OpenAI has launched ChatGPT Agent, its most sophisticated AI agent yet, introducing major upgrades that transform ChatGPT from a conversational assistant into an autonomous digital agent capable of handling complex, multi-step tasks across the internet and connected applications. The rollout began on July 17, 2025, and is available initially to ChatGPT Pro, Plus, and Team subscribers, with Enterprise and Education access planned soon.
What Is ChatGPT Agent?
ChatGPT Agent builds on OpenAI’s previous agentic tools like Operator and Deep Research, merging their capabilities to allow ChatGPT to plan, reason, and execute intricate workflows. It operates using a secure, sandboxed virtual computer, enabling it to interact with web browsers, run code in a terminal, edit files, and integrate directly with APIs for apps such as Gmail, Google Calendar, Notion, and GitHub. Unlike basic chatbots, ChatGPT Agent autonomously breaks down complex requests, plans the steps, switches between tools, and completes the task—notifying the user when finished.
Key Capabilities
- Advanced Planning and Execution: Decomposes complex instructions into discrete steps, executes them sequentially, and adapts based on real-time results.
- Integrated Tool Use: Seamlessly switches between browsers, terminals, file systems, and code environments to accomplish varied tasks.
- Web & API Automation: Can browse the web, fill online forms, download and analyze files, interact with external APIs, and summarize or extract information.
- Document & Report Generation: Creates slide decks, fills out spreadsheets, analyzes CSV files, and delivers complete reports end-to-end.
- File System & Terminal Access: Navigates folders, edits documents, writes and executes scripts, and runs commands in a controlled environment.
- App Connectivity: Automates workflows across platforms like Google Workspace and GitHub via secure connectors.
User Controls, Safety, and Privacy
OpenAI has implemented several layers of safety and control:
- User Approval: All irreversible actions (e.g., submitting forms, making purchases) require explicit user approval.
- Sandboxed Browsing: Virtual environment ensures the agent never has direct access to passwords or sensitive credentials.
- Activity Controls: Users can pause, cancel, or intervene in the agent’s activity at any step; all sessions are auditable, and a complete task history is maintained.
- No Persistent Memory: For added privacy, ChatGPT Agent does not retain conversation memories between sessions.
- Sensitive Task Monitoring: Real-time filters prevent misuse in domains like biology, chemistry, and cybersecurity.
Performance Benchmarks
ChatGPT Agent sets new marks in AI task automation:
- HumanEval: 41.6% pass@1, up from 35.7%, reflecting better code generation and reasoning.
- FrontierMath: 27.4% accuracy, the highest among available public models.
- Leads in AutoDemos, SimulEval, and Multi-hop Retrieval for tasks involving data analysis and spreadsheet workflows.
Limitations
- Generated slide decks may lack polish.
- Spreadsheets might have occasional formatting or accuracy issues.
- The agent may need clarification with ambiguous requests, and switching between tools can sometimes cause delays.
- Full automation is not allowed for critical actions—these always require user intervention.
Availability & Access
- Available now for Pro, Plus, and Team subscribers.
- Pro users receive 400 agent messages monthly; Plus and Team users get 40 free agent messages each month, with top-ups available.
- Rollouts for Enterprise and Education plans are expected later in the summer.
- The older Operator beta is being discontinued; Deep Research mode remains as an optional, slower, more detailed research tool.
- To activate, users select “Agent mode” from ChatGPT’s list of tools.
With this launch, OpenAI is positioning ChatGPT Agent as a true next-generation digital assistant—capable of not just responding to questions, but autonomously acting on behalf of users for a wide range of business and productivity scenarios.