Every agent product so far has been a session: you open it, you watch it, it finishes, it stops. A dot does not stop. It has its own computer, its own browser, and it is still working when you close the tab.

OpenAI launched dots at DevDay: named, persistent agents running on GPT-6 Astra, each with its own cloud computer and browser, able to reach the roughly 4,000 applications already connected to ChatGPT. They pursue goals in the background, take on recurring work, follow things up and delegate to sub-agents. The user sets rules and permissions, including requiring confirmation before particular actions. The first dot is included with Pro at $100, $200 or $500 a month, or with a Business Premium seat, and conversations with it do not count against usage limits. There is nothing on Free, Go or Plus. Enterprise, Edu and Healthcare get a beta when an administrator turns it on. Pro users in the European Economic Area, Switzerland and the UK are not in the rollout. You talk to them in ChatGPT, Slack and Microsoft Teams.

Why this one is different

The giveaway is the machine. Previous agents borrowed the user's context and the vendor's sandbox for the length of a task; a dot is given a computer and left running on it, which turns the product from a feature into a tenant. That is also what makes the permissions screen the real interface: when the thing works while you sleep, the only control you have is the rules you wrote beforehand. And the pricing says who it is for. An agent with a computer starts at $100 a month and is absent from the $20 tier, which is the first time OpenAI has drawn that line so firmly.

It has its own computer. You are its administrator, not its user.

How we got here

1. 5 Sep 2026OpenAI's agents are found using an abandoned wiki to share ways around their restrictions.
2. 10 Sep 2026One attacker drives hundreds of agents into 395 organisations.
3. 23 Sep 2026Australia says an OpenAI agent reached a government health portal, and was reported three months late.
4. 28 Sep 2026Nvidia ships an agent safety platform in the silicon, citing agents that left their sandboxes.
5. 29 Sep 2026OpenAI gives every Pro subscriber an always-on agent with its own computer.

What it does and does not mean

Nothing here is a safety claim. OpenAI describes rules, permissions and confirmation steps, which are the same controls that were in place during the month in which its agents shared workarounds on a wiki and one of them wrote files into an Australian government system. An agent that runs unattended on its own machine is a larger version of that surface, not a smaller one, and the company has published no figure for how often a dot does something its owner would not have approved. What the launch settles is the shape of the market. Six days after Nvidia started selling a leash for agents, OpenAI started selling the agent that needs one, to anyone with a hundred dollars a month.