The AI that thinks, acts and automates via its own “computer”
San Francisco, July 17, 2025 – OpenAI launched a major update to ChatGPT: a new “agent mode” called ChatGPT Agent, which can proactively complete complex tasks using its own virtual browser, file‑generation tools, programming terminals, and API connectors.
A unified agentic model
Built by combining last year’s Operator (AI-controlled browser) and Deep Research tools, ChatGPT Agent merges both into one powerful system. It can autonomously:
- Navigate websites to fill forms or place orders.
- Conduct deep multi‑page research for reports.
- Generate files like PowerPoint decks and Excel spreadsheets.
- Run code in a terminal.
- Call APIs (e.g., Google Drive, SharePoint) for data access.
- Switch flexibly between visual and text‑based browsing modes.
Who can access the ChatGPT AI Agent and how does it work?
ChatGPT Agent is available starting July 17, 2025 to ChatGPT Pro, Plus, and Team subscribers via the “agent mode” dropdown in the chat tool. Pro users will get access by the end of day, while Plus and Team users will get access over the next few days. Enterprise and Edu users will get access in the coming weeks. Usage is capped at 400 prompts/month for Pro users and 40/month for Plus and Team. Enterprise and Education tiers follow later this summer; free-tier access is TBD. As per usual the roll out dates should more so be interpreted as their goal, and not a given as in the past some rollouts also took several weeks longer than expected.
Practical use cases, from cupcakes to slide decks
In demonstrations, the AI agent has:
- Ordered dozens of cupcakes (which took ~1 hour).
- Planned date nights.
- Analysed competitors and generated slide decks on Nvidia’s Q1 earnings (about 25 minutes each).
- Formulated Japanese breakfast shopping lists and purchased ingredients automatically.
Average task durations are estimated at 10–15 minutes, though complex multi-step actions may take longer.
OpenAI Agent safety
OpenAI has implemented robust safeguards:
- A “watch mode” requires user presence during risky tasks (e.g., financial or personal data entry).
- The agent prevents browsing social media and financial platforms.
- ChatGPT’s memory feature is deliberately disabled to avoid potential prompt‑injection misuse.
- The system is classified as “high capability” in the biological/chemical domain, triggering enhanced safety protocols.
Leadership and strategy
The project is led by product lead Yash Kumar, who emphasizes the goal of extending ChatGPT’s reach “beyond the screen into real life,” from calendar management and meal planning to summarizing meetings.
OpenAI frames ChatGPT Agent as a critical step in its efforts to monetize ChatGPT while handling the costs of large‑scale AI operations amid growing competition from Meta, Google, Microsoft’s xAI and others.
Summary
- What it is: A multi-functional AI agent that can think and act through a browser, run code, generate office files, and call APIs.
- Who can use it now: ChatGPT Pro, Plus, Team; broader access coming for Enterprise/Education.
- Key safeguards: Watch mode, restricted browsing, no memory, high-risk content filters.
- Why it matters: Signals OpenAI’s shift from chat assistant to fully automated personal and enterprise AI agent.
For more information go to the official OpenAI Announcement on the ChatGPT AI Agent
