OpenAI is introducing a new general-purpose AI agent in ChatGPT, which the company claims can execute a diverse array of computer-based tasks on behalf of users.
According to OpenAI, the agent is capable of generating modifiable presentations and slideshows, running code, and autonomously navigating a user’s calendar.

The tool, ChatGPT agent, integrates several capabilities from OpenAI’s previous agentic tools, such as Operator’s capacity to navigate websites and Deep Research’s capacity to compile information from numerous websites into a concise research report.
OpenAI asserts that users will be able to communicate with the agent by prompted ChatGPT in natural language.

Subscribers to OpenAI’s Pro, Plus, and Team plans will receive the ChatGPT agent on Thursday. Users may activate the utility by selecting “agent mode” from the selection menu of tools in ChatGPT.
ChatGPT is now capable of performing tasks on its own computer.
This section introduces the ChatGPT agent, a unified agentic system that combines the conversational strengths of ChatGPT, the action-taking remote browser of Operator, and the web synthesis of deep research. https://twitter.com/7uN2Nc6nBQ
— OpenAI (@OpenAI) July 17, 2025
The launch of ChatGPT agent is OpenAI’s most ambitious endeavor to transform ChatGPT into an agentic product that can outsource tasks and take actions for users, rather than merely responding queries.
Numerous AI agents have been introduced by Silicon Valley companies, such as OpenAI, Google, and Perplexity, in recent years, with the intention of accomplishing this.
Nevertheless, these initial AI agents have demonstrated their inability to perform intricate tasks and appear to be less appealing as products than the ultimate vision that technology executives present regarding AI agents.

Nevertheless, OpenAI asserts that the ChatGPT agent is significantly more capable than its previous offerings.

The company’s new agent has the ability to access ChatGPT connectors, which enable users to connect applications such as Gmail and GitHub.
This facilitates the agent’s ability to locate pertinent information in response to your inquiries. According to OpenAI, the ChatGPT agent has the ability to access a terminal and utilize APIs to access specific applications.

OpenAI proposes that users can utilize the ChatGPT agent to “plan and purchase ingredients to prepare a Japanese breakfast for four” and “analyze three competitors and generate a slide deck.”
These capabilities necessitate the ChatGPT agent to parse through websites, devise a course of action, and employ tools- complexities that OpenAI has not previously attempted to address with agents.

OpenAI asserts that the ChatGPT agent’s model provides cutting-edge efficacy on numerous benchmarks.

The ChatGPT agent model, according to the company, achieves a pass rate of 41.6% on Humanity’s Last Exam (pass@1), a challenging examination that encompasses over one hundred subjects and includes thousands of questions.
That is approximately twice as high as the scores of OpenAI’s o3 and o4-mini on the test.

OpenAI asserts that the ChatGPT agent achieves a 27.4% score on FrontierMath, one of the most challenging known math benchmarks, when it is provided with access to tools, including a terminal for code execution.
The previous state-of-the-art score was achieved by o4-mini, which achieved a mere 6.3%.

OpenAI indicates that the ChatGPT agent was created with safety in mind, primarily due to the fact that the product features new capabilities that could render it more dangerous in the hands of a malicious actor.
OpenAI has previously issued a cautionary note regarding the potential for agentic models to exhibit more hazardous capabilities.

In a safety report for the ChatGPT agent, OpenAI has classified the model as “high capability” in the biological and chemical weapon domains.
This designation is defined in OpenAI’s Preparedness Framework as a model that has the potential to “amplify existing pathways to severe harm.”
Although OpenAI does not have direct evidence of this, it has chosen to take a precautionary approach and activate new safeguards to mitigate these risks.

A monitor that operates in real time as users interact with the product is among the new safeguards for ChatGPT agent. According to OpenAI, a classifier is applied to each query inputted into the ChatGPT agent to ascertain whether the request pertains to biology.
If so, OpenAI processes the ChatGPT agent’s response through a secondary monitor to ascertain whether the content has the potential to elicit a biological threat.

Additionally, OpenAI has stated that it has disabled the memory feature of ChatGPT for this agent in order to prevent its misuse. OpenAI’s memory feature enables the chatbot to access information from previous user conversations in other sections of ChatGPT.
Nevertheless, OpenAI asserts that malicious actors may exploit the ChatGPT agent’s feature to exfiltrate sensitive data through prompt injection attacks.
Nevertheless, the organization has indicated that it may reconsider the inclusion of the feature in the future.

The real-world capabilities of the ChatGPT agent are yet to be determined, despite its remarkable sounding capabilities.
Agent technology has been comparatively fragile in its interaction with the real world up to this point. Nevertheless, OpenAI asserts that it has created a more advanced model that is capable of fulfilling the potential of AI agents.