OpenAI has announced the launch of a new general-purpose tool within ChatGPT, called the ChatGPT agent.
This is a major step towards making AI more action-oriented. Rather than simply answering questions, the agent performs real tasks on behalf of the user.
Available starting July 18, 2025, the ChatGPT agent is being rolled out to users on the Pro, Plus, and Team plans.
It can be enabled by selecting “agent mode” from the dropdown menu in ChatGPT.
ChatGPT Agent
The ChatGPT agent can complete a range of computer-based tasks prompted by plain language text.
Then, the agent will respond by taking action rather than just offering information. Some of its core abilities include:
- Navigating calendars
- Creating editable presentations and slides
- Running code
- Researching topics across websites
- Connecting with external apps like Gmail and GitHub
It achieves this by combining several of OpenAI’s previous agentic tools. These include Operator, which could interact with websites, and Deep Research, which summarizes data from multiple online sources.
Now, these capabilities are merged into one fully functional agent.
Benchmarks
OpenAI says the underlying model for the ChatGPT agent shows top-level performance in testing.
For example:
- On Humanity’s Last Exam, it scored 41.6% (pass@1). This is double what earlier models o3 and o4-mini achieved.
- On FrontierMath, it scored 27.4% when using tools. In comparison, o4-mini scored just 6.3%.
These benchmarks suggest improved reasoning and reliability, especially when handling multi-step tasks.
Also read: Humanity’s Last Exam: The One Test AI Couldn’t Beat
Real-Life Examples
The agent is designed for a wide range of users. OpenAI shared several use cases:
- Planning a Japanese breakfast for four people
- Analyzing three competitors
- Creating a complete slide deck from the results
In these examples, the agent reads data online, develops a plan, and uses multiple tools, all based on a single prompt.
Integrations
The ChatGPT agent can access external services through ChatGPT Connectors. Users can connect their Gmail, GitHub, and other tools.
This allows the agent to retrieve specific information and act on it. It also has access to a code execution terminal, external APIs, and file and web browsing tools
This combination makes the agent act as a true virtual assistant.
Safety Measures
OpenAI has stated that the new agent introduces safety concerns, especially due to its advanced capabilities.
As a result, the company has implemented strict safeguards. Key protections include:
- Real-time monitoring: Each prompt is scanned for sensitive content, especially those related to biology. If flagged, an additional layer of review is triggered.
- Memory is disabled: Unlike other versions of ChatGPT, this agent does not recall past chats. This helps prevent malicious prompt injection.
- Risk awareness: OpenAI has flagged the model as “high capability” in the biological and chemical weapons domains. Although there is no evidence of misuse, the company has taken a cautious stance.
Availability
The ChatGPT agent is now live for ChatGPT subscribers on the Pro, Plus, and Team tiers. It can be activated with a simple toggle.
OpenAI has not announced when or if it will expand access to free users. However, this release offers a strong reason to upgrade.
