OpenAI’s dots work 24/7 on their own cloud computer

Dots, the always-on agents OpenAI unveiled at its DevDay on September 29, each get a cloud computer of their own and a rulebook you write: what the agent may do on its own, what waits for your approval, and what it must never touch. They run on GPT-6 Astra and are rolling out in ChatGPT to Pro and Business Premium users in eligible markets, OpenAI announced on X and in a longer post on its site. Enterprise workspaces, including Edu and Healthcare, get a beta that an admin has to switch on.

The main difference from a chat assistant is that a dot keeps working between conversations. It has its own browser and machine and connects to more than 4,000 apps through OpenAI’s plugin ecosystem. It can also run several projects at once without you juggling separate threads. You reach it in ChatGPT on web, desktop and mobile, in Slack or Teams, or on a voice call, with texting “coming soon,” and it carries the same context across all of them. It also contacts you first when it has progress, a question or a decision that needs you. You can open the dot’s computer at any time to inspect its work. Your own laptop stays separate unless you give the dot permission to connect to it.

This is OpenAI’s third attempt at an agent that uses a computer for you. Operator, announced on January 23, 2025, read screenshots and drove a browser with mouse and keyboard input. ChatGPT agent folded that into ChatGPT on July 17, 2025, and the standalone Operator app shut down on August 31 that year. ChatGPT agent was itself removed from ChatGPT in early August 2026. Those agents ran a task when you asked for one. A dot works toward standing goals around the clock and, OpenAI says, will “learn from feedback over time.”

The background work has its own name. When you aren’t actively working with it, a dot does what OpenAI calls “proactive research”: it looks through the apps you’ve already connected for ways to help, using tools restricted to read-only. In that mode it can’t send messages, change app content, or control your browser or computer. OpenAI’s one outside example shows how the handoff works. An early tester’s dot noticed he had forgotten to invoice a publication and prepared the invoice. It sent the invoice only after he approved.

Once a dot moves from reading to acting, three layers apply. Built-in rules set the default for when it acts independently and when it asks. Custom Rules, managed through your existing ChatGPT app controls, let you allow specific actions, require approval for them or block them outright. On top of both, auto-review checks any action that could affect your accounts or share information against your instructions, your rules and OpenAI’s safety requirements, and decides what proceeds, what needs approval and what you have to do yourself. Some sensitive tasks, such as changing a password, always stay with you. Dots can use saved passwords to sign into supported websites without the model seeing them, and Activity View lets you follow everything, background work included, and redirect the dot.

OpenAI says its security safeguards “help defend against malicious instructions” and watch for potentially harmful behavior, and a monitoring system can pause or stop a dot when it detects a safety concern. The verb is only “help,” and with reason: OpenAI’s own red-teaming has produced prompt injections that copy themselves from one agent reply to the next.

Training rules split by plan. Content from ChatGPT Business, Enterprise and Edu workspaces isn’t used to improve OpenAI’s models by default. On personal plans you decide whether your dot’s conversations and work are used. OpenAI says it doesn’t train directly on proactive research or on a dot’s notes to itself, though information from them may be used to inform an eligible conversation or task, depending on your settings.

Your first dot comes with a Pro or Business Premium plan at no extra cost, along with an allowance for deeper work and extended limits for the first month after launch. Talking to the dot doesn’t count toward ChatGPT usage limits. Tasks it starts or manages in Codex or ChatGPT Work do count as usual, and on the $200 Pro plan for new subscribers that pool shrank at the same DevDay, from 20 times Plus usage to 10 times. OpenAI plans to let you add more dots later, and to scale each one by making it faster or raising how much work it can take on per month.

Inside OpenAI, the company says, dots start investigating a bug as soon as it appears in Slack, turn a new design into a working app while the team works through customer feedback, and keep a planning cycle on track toward its deadline. That internal testing also covered procurement, invoice processing, email marketing, customer support and commercial contracting. That is the ground for specialist dots, which OpenAI is previewing for organizations. Each one gets its own identity for access management, IT-provisioned hardware and deep integrations with a company’s systems of record, so it can own a well-defined job instead of working for one person. They start as focused enterprise pilots, with OpenAI engineers working directly with each organization to define the dot’s responsibilities, its tools and how people review its work. OpenAI is also working with Microsoft to bring specialist dots under the governance and security controls of Microsoft Agent 365, so businesses can manage them with Microsoft tools they already use.

Setup starts on a desktop: you create the first dot in the ChatGPT desktop app or a desktop browser, connect your apps, and “let it introduce itself.” After that you can message it from the ChatGPT mobile app. OpenAI’s own caveat comes with the launch: “Dots can still make mistakes, so always review consequential work.”