OpenAI Unveils All-Purpose Chatbot: Meet the General Purpose Agent in ChatGPT
In an unprecedented development, OpenAI is set to introduce a new general-purpose AI agent within its platform, ChatGPT. The company emphasizes that this advanced AI agent will be capable of executing a wide array of computer-based tasks on behalf of users, encompassing calendar management, generation of editable presentations and slideshows, and code execution.
Dubbed the ChatGPT Agent, this tool is a fusion of various capabilities from OpenAI’s previous agentic tools, incorporating the functionality of Operator for website navigation and Deep Research for synthesizing information from multiple sources into concise research reports. Users will interact with the agent by prompting ChatGPT in natural language.
Commencing on Thursday, OpenAI will make the ChatGPT Agent available to subscribers of its Pro, Plus, and Team plans. To initiate the tool, users can simply select “agent mode” from the dropdown menu of tools within ChatGPT.
This launch signifies OpenAI’s most ambitious endeavor to transform ChatGPT into an agentic product capable of taking action and offloading tasks for users, beyond merely answering questions. Over the past few years, numerous AI agents from tech giants like OpenAI, Google, and Perplexity have been unveiled with similar promises. However, early versions of these AI agents have grappled with complex tasks, often falling short of the compelling products envisioned by tech executives.
Despite these historical challenges, OpenAI asserts that the ChatGPT Agent surpasses its previous offerings in terms of capability. The new agent can access ChatGPT connectors, enabling users to link apps such as Gmail and GitHub, thereby allowing the agent to retrieve relevant information for user prompts. Furthermore, OpenAI claims that the ChatGPT Agent has access to a terminal and can utilize APIs to interact with specific applications.
The model powering ChatGPT Agent demonstrates state-of-the-art performance on various benchmarks, according to OpenAI. On Humanity’s Last Exam (pass@1), a notoriously challenging test consisting of thousands of questions across over one hundred subjects, the ChatGPT Agent scored 41.6%, marking approximately double the score achieved by OpenAI’s o3 and o4-mini on the same test.
On FrontierMath, one of the most difficult known math benchmarks, OpenAI reports that the ChatGPT Agent scores 27.4% when equipped with tools such as a terminal for code execution. The previous state-of-the-art score was recorded by o4-mini, which managed only 6.3%.
OpenAI acknowledges that developing the ChatGPT Agent with safety in mind was crucial due to its novel capabilities, which could potentially pose new risks in the hands of malicious actors. However, the true extent of the ChatGPT Agent’s capabilities remains to be determined.
