OpenAI Unveils ChatGPT Agent Capable of Performing Tasks Independently

OpenAI has launched a new universal AI agent within ChatGPT, capable of executing a wide range of computer tasks on behalf of the user.

ChatGPT can now do work for you using its own computer.

Introducing ChatGPT agent—a unified agentic system combining Operator’s action-taking remote browser, deep research’s web synthesis, and ChatGPT’s conversational strengths. pic.twitter.com/7uN2Nc6nBQ

— OpenAI (@OpenAI) July 17, 2025

The company claims it can automatically manage a user’s calendar, create editable presentations and slides, and even run code.

The ChatGPT agent integrates several features from previous agent solutions, including Operator’s ability to click through websites and Deep Research’s capability to gather information from numerous sites and provide a concise analytical report.

Users can interact with the tool in natural language through dialogue with the chatbot.

Initially, the AI agent is available to Pro, Plus, and Team subscribers. To activate it, one must select “agent mode” from the ChatGPT tools dropdown menu.

OpenAI asserts that the new ChatGPT agent significantly surpasses other solutions. It can employ ChatGPT connectors to link applications like Gmail and GitHub to find necessary information and respond to queries. It also has access to a terminal and can use the API.

The digital assistant’s skills include planning and purchasing ingredients for a Japanese breakfast for four and analyzing three competitors followed by preparing a presentation.

Tests

The underlying model of the tool demonstrates advanced results in several benchmarks, according to OpenAI. In Humanity’s Last Exam—a challenging test comprising thousands of questions across more than a hundred subjects—ChatGPT agent scores 41.6%. This is roughly twice the scores of o3 and o4-mini.

In one of the most complex mathematical analyses, FrontierMath, the neural network scored 27.4%. The previous record was held by o4-mini at 6.3%.

Security

The startup noted that special attention was paid to security issues during the development of the ChatGPT agent, given its enhanced capabilities that could cause harm if misused.

In the report, the model is classified as “high capability” in the area of biological and chemical weapons. This indicates it could amplify existing pathways for causing significant harm. However, OpenAI emphasizes that there is no direct evidence of such a threat but applies a preventive approach and introduces additional security measures. These include:

• an online monitoring module—all user requests pass through a classifier that determines if the inquiry is related to biological topics. If so, the response is further checked by a second mechanism for potential threats;
• disabling the memory function—this is done to prevent data leaks through malicious prompt injection attacks.

In July, it was revealed that OpenAI revised its security system to protect intellectual property from corporate espionage amid concerns about theft by Chinese competitors.

Previously, ChatGPT was trained to connect to more internal sources and obtain contextual information in real time.