OpenAI has launched a new general-purpose AI agent within ChatGPT, designed to automate a wide array of computer-based tasks for users. This advanced tool integrates functionalities from previous OpenAI agentic tools.
The ChatGPT agent is being rolled out on Thursday to subscribers of OpenAI’s Pro, Plus, and Team plans. Users can activate this feature by selecting “agent mode” from ChatGPT’s dropdown menu. This launch signifies OpenAI’s most ambitious effort to transform ChatGPT from a question-answering system into an actionable, task-offloading agent. According to OpenAI, the ChatGPT agent combines Operator’s action-taking remote browser, deep research’s web synthesis, and ChatGPT’s conversational strengths.
Unlike earlier AI agents that often struggled with complex tasks, OpenAI asserts that ChatGPT agent is significantly more capable. It can access ChatGPT connectors, allowing integration with applications such as Gmail and GitHub, enabling it to retrieve relevant information for prompts. The agent also possesses access to a terminal and can utilize APIs to interact with various applications. OpenAI highlights practical applications for ChatGPT agent, such as planning and purchasing ingredients for a meal or analyzing competitors to create a slide deck. These tasks demand the agent to parse websites, formulate action plans, and use tools, representing a leap in complexity compared to previous agent capabilities.
The underlying model powering ChatGPT agent demonstrates state-of-the-art performance on several benchmarks. On “Humanity’s Last Exam” (pass@1), a rigorous test comprising thousands of questions across over a hundred subjects, the model achieved a score of 41.6%. This is roughly double the scores of OpenAI’s o3 and o4-mini models. For “FrontierMath,” a challenging math benchmark, ChatGPT agent scored 27.4% when provided with tools like a terminal for code execution, significantly outperforming the previous state-of-the-art score of 6.3% from o4-mini.
OpenAI has emphasized safety considerations in the development of ChatGPT agent, acknowledging the heightened risks posed by its advanced capabilities. In a safety report for ChatGPT agent, OpenAI classified the model as “high capability” in biological and chemical weapon domains. This classification indicates the model’s potential to “amplify existing pathways to severe harm,” even though direct evidence is not yet available. Consequently, OpenAI has implemented new safeguards as a precautionary measure.
Key safety measures include a real-time monitoring system. A classifier analyzes every prompt entered into ChatGPT agent to detect biology-related requests. If identified, the agent’s response is then passed through a second monitor to determine if the content could contribute to a biological threat. Furthermore, OpenAI has temporarily disabled ChatGPT’s memory feature for this agent to prevent potential misuse, such as sensitive data exfiltration through prompt injection attacks.
Despite the impressive capabilities outlined by OpenAI, the true real-world performance of ChatGPT agent remains to be seen. Historically, agent technology has demonstrated fragility when interacting with complex real-world scenarios. Nonetheless, OpenAI expresses confidence that its more capable model can deliver on the long-promised vision of AI agents.
