OpenAI has unveiled a major update to ChatGPT, introducing Agent Mode — a new feature that empowers the AI to autonomously execute multi-step, real-world tasks on your behalf. With its own secure virtual computer, ChatGPT can now handle requests such as analyzing your schedule, researching competitors, planning meals, and even assembling editable presentations and spreadsheets— all from start to finish using instructions you provide.
Seamless Integration of Advanced AI Systems
At the heart of Agent Mode is a unified system merging the strengths of Operator (web page interaction), Deep Research (information synthesis), and ChatGPT’s conversational reasoning. This lets ChatGPT pivot smoothly between research, reasoning, and action—navigating sites, filtering results, logging in securely, executing code, and delivering polished outputs—without breaking your workflow.
Complete User Control
Users retain full control throughout the process. ChatGPT always seeks confirmation before taking consequential actions, and you can interrupt, adjust, or halt tasks at any moment. This collaborative design ensures transparency and keeps you in charge, even as AI does more of the heavy lifting.
How to Get Started With Agent Mode?
Starting immediately, Pro, Plus, and Team users can access Agent Mode via the tools dropdown in the message composer. Wider rollout to Enterprise and Education customers will follow in the coming weeks. Simply select 'agent mode,' describe your desired outcome, and watch as ChatGPT executes each step with real-time narration and easy options to intervene. Recurring tasks like weekly reports can also be scheduled, streamlining your workload.
Expanded Toolset and Workflow Integration
- Visual and Text-Based Browsers: Navigate websites and efficiently process online information.
- Terminal and API Integration: Run code or connect external services directly from ChatGPT.
- Connectors: Seamlessly link Gmail, GitHub, calendars, and more, so ChatGPT can access relevant information as needed.
You can grant controlled access for deeper research or instant task execution, always retaining the option to step in for added security.
Real-World Performance: Outperforming Past Models
ChatGPT Agent shows marked improvement over previous versions on industry-standard benchmarks:
| Task/Benchmark | OpenAI o3 | Deep Research | Human | ChatGPT Agent |
|---|---|---|---|---|
| Humanity's Last Exam | 20.3–26.6% | 26.6% | - | 41.6% |
| FrontierMath | 10.3–19.3% | - | - | 27.4% |
| DSBench (Data Analysis) | 34.1% | - | 64.1% | 89.9% |
| SpreadsheetBench | 16.8–23.3% | - | 71.3% | 45.5% |
In data science and professional tasks—like investment banking modeling or spreadsheet editing—ChatGPT Agent outperformed both earlier models and human participants in several tests, highlighting its powerful real-world utility.
Enhanced Security and Privacy
- Strengthened controls against unauthorized actions and prompt injection attacks.
- New privacy features: instant browser data deletion, secure browser takeover.
- Mandatory user confirmations for sensitive actions.
- 'Watch Mode' for oversight on critical tasks like sending emails.
- Enhanced safeguards for biological and chemical knowledge use.
OpenAI’s safety framework for Agent Mode has been thoroughly reviewed by external experts, including biosecurity researchers, and will continue to evolve with advancing AI capabilities.
ChatGPT Agent Availability
ChatGPT Agent is rolling out now. Pro users will get it by today, and Plus and Team users will see it in the next few days. Enterprise and Education users will get it in the coming weeks.
- Message limits: Pro users get 400 messages/month. Others get 40 messages, with extra available via credits.
- Deep research: Now built into the Agent. You can still pick the original "Deep Research" mode if you want longer, more detailed answers.
Note. It’s not yet available in the European Economic Area or Switzerland.
The Operator preview site will stay live for a few more weeks, then shut down.
Join the conversation! Your thoughts help the community grow.