Openai is launching a new general-purpose AI agent with ChatGPT. The company says it can complete a variety of computer-based tasks on behalf of its users. According to Openai, agents can automatically navigate the user’s calendar, generate editable presentations and slideshows, and run code.
This tool, known as the ChatGpt Agent, combines some of the features of Openai’s previous agent tools. This includes the ability to click on the operator’s website and the ability of Deep Research to synthesize information from many websites into concise research reports. Openai says users can interact with agents simply by prompting ChatGPT in natural language.
ChatGpt agents will be rolling out to Openai’s Pro, Plus and Team Plan subscribers on Thursday. To activate the tool, users can select “Agent Mode” in the Tools drop-down menu in ChatGPT.
The launch of the ChATGPT agent represents Openai’s boldest attempt to transform it into an agent product that can perform actions and off-road tasks to users rather than answer questions. In recent years, Silicon Valley companies, including Openai, Google and Perplexity, have announced numerous AI agents that have committed to doing just that. However, these early versions of AI agents have proven to wrestlech with complex tasks, and seem less convincing as a product than the ultimate Vision Tech executive pitches around AI agents.
That said, Openai says the ChatGpt agent is much more capable than its previous products.
The company’s new agents can access the ChatGpt connector, allowing users to connect apps such as Gmail and GitHub, allowing agents to find relevant information at the prompt. According to Openai, ChatGpt agents can access devices and use APIs to access certain apps.
Openai suggests that users can tap on the CHATGPT agent to “plan and buy ingredients to make four Japanese breakfasts” and “analyze three competitors and create a slide deck.” These types of functionality require the ChatGPT agent to parse the website, plan a set of actions, and use tools. This is a much more complicated task than Openai previously worked with agents.
TechCrunch Events
San Francisco
|
October 27th-29th, 2025
According to Openai, the underlying ChatGpt agent model offers cutting-edge performance on several benchmarks.
The company says the ChatGPT agent model won 41.6% in the final test for humanity (pass@1). This almost doubles what Openai’s O3 and O4-Mini scored in the Test.
One of the most challenging mathematics benchmarks, Frontiermath says that if the ChatGPT agent has access to a tool such as a terminal for code execution, the ChatGPT agent scored 27.4%. Previous cutting-edge scores come from the O4-Mini, which scored just 6.3%.
Openai notes that it has developed the ChatGpt agent with safety in mind, primarily to present new features that could make the product even more dangerous in the hands of bad actors. Openai has previously warned that the agent model could potentially present more dangerous features.
In its CHATGPT Agent safety report, Openai says it designates the model as “high capacity” in the biological and chemical weapons domains. This is defined as a model with the ability to “amplify existing pathways to harm” in Openai’s preparation framework. Although Openai notes that there is no direct evidence of this, it has decided to take a preventive approach and activate new safety measures to mitigate these risks.
The new Safeguard for CHATGPT Agent includes a monitor that works in real time when users interact with the product. Openai says it will run the classifier on every prompt entered into the CHATGPT agent to determine if the request is biology-related. In that case, Openai will perform the CHATGPT agent response on the second monitor. This determines whether content can be used to invoke biological threats.
Openai also says that the agent has disabled the memory feature in ChatGpt to prevent misuse. In other parts of ChatGPT, OpenAI’s memory feature allows chatbots to view information from previous user chats. However, Openai says bad actors can use the ChatGPT agent’s capabilities to remove sensitive data through rapid injection attacks. The company says it could revisit the addition of features in the future.
ChatGpt’s agents sound impressive, but it’s still not clear how capable they are in the real world. Until now, agent technology has proven relatively vulnerable when interacting with the real world. That said, Openai says it has developed a more capable model that can fulfill the promise of AI agents.
This story has been updated with more information.
