OpenAI has introduced ChatGPT Agent, a general-purpose AI agent that can control a virtual computer to complete multi-step tasks for users. The system unifies the company’s earlier Operator and Deep Research tools, allowing ChatGPT to browse the web, run code, fill out forms, generate slideshows and spreadsheets, and interact with services such as Gmail and GitHub. The feature, available through a new “agent mode,” begins rolling out immediately to ChatGPT Pro, Plus and Team subscribers, with Enterprise and Education customers to follow. Pro users receive 400 agent messages per month, while Plus users get 40, and additional usage can be bought with credits. OpenAI says ChatGPT Agent requests permission before taking consequential actions and activates a “Watch Mode” for sensitive sites such as banking and email. The model is classified as high-risk for biological and chemical misuse under the company’s Preparedness Framework, triggering always-on monitoring and other safeguards developed after extensive red-team testing. Initial benchmarks show the dedicated model powering the agent scoring 41.6 on Humanity’s Last Exam, 45.5 percent on SpreadsheetBench and outperforming human baselines on roughly half of OpenAI’s internal knowledge-work evaluations. The launch positions OpenAI to compete in the fast-growing market for autonomous AI assistants capable of handling real-world workflows.