OpenAI just gave ChatGPT a major upgrade, and it fundamentally changes what the chatbot can do. Forget just generating text. Starting today, Pro, Plus, and Team subscribers can put ChatGPT to work on complex, multi-step tasks by giving it control of a virtual computer. Think of it as the chat interface finally growing arms and legs. By typing a prompt like “analyze three competitors and create a slide deck,” the model will now proactively choose from a toolkit—a visual web browser, a text-based browser, a terminal, and direct APIs—to click around the web, gather data, run code, and compile an editable deliverable, all without you having to micromanage each step.
This isn’t just a single model working in isolation. OpenAI has stitched together the previously separate strengths of its Operator (which could interact with web GUIs) and Deep Research (which synthesized information) into one unified agentic system. The company admitted that the old approach was siloed, noting that many tasks users tried to hand to Operator were actually a better fit for deep research. Now, the agent shifts fluidly between reasoning and action, opening a page in a text-based browser to reason over a large document, then switching to a visual browser to click a button behind a login wall. You can even connect apps like Gmail or GitHub through ChatGPT connectors to give it a direct line to your data.
The control scheme is designed to feel collaborative rather than like handing over the keys. The agent asks for permission before any consequential action, and you can interrupt it mid-flow to refine the task, take manual control of the browser to log in, or stop the process entirely and grab whatever partial results it has gathered. If a task is running long, you can ask for a status update, and the mobile app will ping you when the job is done. The virtual computer preserves all session context, so the model can download a file, manipulate it in the terminal, and display the output visually without getting lost.
OpenAI is positioning this as the first iteration. The blog post promises “significant improvements regularly,” suggesting we’re seeing the floor, not the ceiling, of this agentic behavior. For now, the ability to seamlessly pivot from a casual chat to delegating a 20-minute research slog is a tangible step beyond the static chatbot paradigm. It turns ChatGPT from a brilliant conversationalist you ask for advice into an assistant you can tell to do the boring stuff. Whether it’s planning a Japanese breakfast with a specific shopping list or briefing you on clients by cross-referencing your calendar with recent news, the real test will be how gracefully it handles the messy, unpredictable reality of the open web without getting stuck in an infinite loop.
