OpenAI Unveils Multi-functional Agent in ChatGPT

OpenAI is unveiling a versatile AI agent via ChatGPT, which the organization asserts can undertake a wide variety of computer-related tasks for its users. This agent can autonomously handle calendar management, create editable presentations and slideshows, and execute code.

Dubbed the ChatGPT agent, this tool consolidates several features from OpenAI’s previous agent-oriented tools. It utilizes Operator’s ability to browse websites and Deep Research’s skill in summarizing information from various sources into a succinct research report. OpenAI mentions that users will engage with the agent by inputting natural language prompts in ChatGPT.

On Thursday, OpenAI will roll out the ChatGPT agent for subscribers of its Pro, Plus, and Team plans. Users can enable this feature by selecting “agent mode” from the dropdown menu in ChatGPT.

The launch of the ChatGPT agent represents OpenAI’s most ambitious initiative to evolve ChatGPT into a practical tool that not only answers questions but also executes tasks for its users. In recent times, numerous companies in Silicon Valley, including OpenAI, Google, and Perplexity, have introduced various AI agents that promise to undertake similar responsibilities. However, prior iterations of these AI agents have had difficulties with complexity and have yet to meet the high expectations set by industry leaders.

Nevertheless, OpenAI contends that the ChatGPT agent surpasses its earlier tools in terms of capability.

The new agent is equipped to use ChatGPT connectors, allowing users to link applications like Gmail and GitHub, enabling the agent to extract pertinent data based on prompts. Moreover, the ChatGPT agent can access a terminal and utilize APIs to interact with specific applications.

OpenAI indicates that users could instruct the ChatGPT agent to “organize and buy ingredients for a Japanese breakfast for four” or “analyze three competitors and create a slide deck.” These tasks necessitate the ChatGPT agent to navigate websites, strategize, and employ various tools, showcasing a higher degree of complexity than OpenAI’s previous attempts with agents.

According to OpenAI, the underlying model of the ChatGPT agent offers top-tier performance on various assessments.

The company reports that the ChatGPT agent model scores 41.6% on Humanity’s Last Exam (pass@1), a rigorous test comprising thousands of questions across multiple subjects. This performance is nearly double that of OpenAI’s o3 and o4-mini models on the same assessment.

On FrontierMath, recognized as one of the most challenging math evaluations, OpenAI states that the ChatGPT agent attains 27.4% when granted resources like a terminal for code execution, significantly exceeding the previous highest score of 6.3% achieved by the o4-mini.

OpenAI underscores safety in the development of the ChatGPT agent, particularly as its new features may present risks if exploited by malicious individuals. The company has previously warned that agentic models might carry more dangerous functionalities.

In its safety report for the ChatGPT agent, OpenAI characterizes the model as “high capability” in areas related to biological and chemical weapons, following its Preparedness Framework, which defines such a model as having the potential to “amplify existing pathways to severe harm.” Although OpenAI lacks direct proof of this capability, it has opted for caution by instituting new safeguards to minimize possible risks.

Among the enhanced protections for the ChatGPT agent is a real-time monitoring system that activates during user interactions. OpenAI claims it acts as a classifier for every prompt submitted into the ChatGPT agent, assessing whether the request pertains to biological matters. If it does, the agent’s response undergoes additional scrutiny through a secondary monitor to determine if the content might pose a biological threat.

While the ChatGPT agent appears commendable on paper, its practical effectiveness is yet to be assessed. Historically, agent technology has exhibited vulnerabilities in real-world applications. Nonetheless, OpenAI is optimistic that it has developed a more competent model capable of fulfilling the ambitions associated with AI agents.