OpenAI has launched its most powerful AI system to date – the ChatGPT Agent, capable of planning, reasoning, and performing complex multi-step tasks using integrated tools like web browsing, terminal access, and API connectivity. Available from 17th July 2025, the agent is rolling out to ChatGPT Pro, Plus, and Team users, marking a major leap in personal AI assistants.
The ChatGPT Agent builds upon the standard ChatGPT experience, evolving it into an autonomous AI assistant capable of executing instructions across different applications and platforms. Instead of just responding to queries, the agent can:
- Break down complex instructions into steps
- Choose the right tools (browser, terminal, file system, etc.)
- Execute tasks independently while keeping users in the loop
- Notify users once tasks are complete
- All actions run within a sandboxed virtual environment, ensuring safety and user control.
ChatGPT Agent – Key Features and Capabilities
- Tool Switching: Accesses browser, terminal, code interpreter, and file systems to complete tasks
- Web Automation: Searches online, logs into services (with permission), downloads data, and summarizes results
- App Integration: Works with Gmail, Google Calendar, Notion, GitHub, and more
- Document Handling: Creates slide decks, edits files, runs scripts, and handles spreadsheets
- Multi-Step Planning: Efficiently executes long workflows across research, coding, and productivity tools
Safety and User Control
- OpenAI has placed a strong emphasis on privacy and safety:
- Users review any irreversible actions, like purchases or form submissions
- All interactions occur in a secure, sandboxed system
- Users can pause, review, or cancel any session
- Full task history is available for transparency
- Restrictions are in place to prevent misuse in sensitive domains (e.g., biosecurity)
Performance and Benchmarks
- The ChatGPT Agent has shown significant improvements across key benchmarks:
- 41.6% pass@1 on HumanEval (vs 35.7% before)
- 27.4% accuracy on FrontierMath, the best among public models
- Leading scores in multi-hop reasoning, data handling, and spreadsheet automation
- These benchmarks validate the agent’s ability to handle real-world, professional-grade tasks.
Limitations
- Despite its impressive performance, the ChatGPT Agent isn’t perfect. Known issues include:
- Basic visual formatting for slides
- Occasional spreadsheet inaccuracies
- Misinterpretation of vague or layered prompts
- Occasional lag in tool switching
- Final decisions (e.g., purchases, critical submissions) still require user approval
OpenAI has committed to improving formatting, tool coordination, and reliability in future updates.
The ChatGPT Agent rolls out starting 17th July 2025, with the following access tiers:
- Pro users: 400 messages/month
- Plus & Team users: 40 messages/month (top-ups available)
- Enterprise & Education: Coming soon
- Replaces the earlier Operator beta
- Deep Research is now fully integrated with ChatGPT
Users will be notified when the agent activates or completes a task, and settings allow full control of its activity.
OpenAI is reportedly working on native support for Microsoft Excel (.xlsx) and PowerPoint (.pptx) file formats within ChatGPT. While users can currently generate such files via text and download them, the new upgrade could allow real-time editing and interaction directly inside ChatGPT’s interface, including Canvas mode.
