On 29 September, at its annual developer conference, OpenAI launched Dots: AI agents that keep working when nobody is at the screen. Each one gets its own cloud computer and browser, runs on the GPT-6 Astra model, and works on the goals it has been given around the clock.
The difference from a chat is persistence. In a chat you ask and get an answer. An agent gets an open-ended job, such as following up on customer enquiries or drafting social posts from every new recording, and keeps at it on its own, learning from the feedback it gets along the way.
An agent with its own computer
A dot connects to more than 4,000 apps through ChatGPT's plugin system. You talk to it inside ChatGPT, in Slack or in Teams, and OpenAI says text messaging is coming soon. It is created in the desktop app or the browser and can then be managed from the phone.
The examples OpenAI showed map closely onto small-business work. On the sales side, the agent compares a prospect's requirements against what the business offers and builds a first proposal. On the content side, it goes through interviews and recordings, finds clips, prepares notes and drafts posts. In one demo it generates invoices on its own and asks for approval before they go out.
Plans and prices
Dots is available from launch day to Pro subscribers and to Business Premium seats on ChatGPT's business plan. OpenAI says Business Premium users get it in every region where ChatGPT is supported. On enterprise plans it only switches on once a workspace admin enables it.
The first dot comes with the subscription at no extra cost. More dots will be sold separately, and OpenAI has not published what they will cost.
A Business Premium seat costs $125 a month billed monthly, or $100 a month billed annually. A standard seat on the same plan costs $25 monthly or $20 annually. Both kinds can sit in one business workspace, so a three-person business can upgrade only the seat of whoever runs the agent. A Premium seat also gives five times the usage of a standard seat and drops the five-hour usage limit.
Rules before access
An agent connected to the inbox, the calendar and the invoicing system can do real damage when it gets something wrong. OpenAI built in several brakes, and they are worth knowing before connecting anything.
By default, the research a dot does in the background is read-only: it gathers information but sends no messages and changes no data. Every other action gets its own rule. The agent can be allowed to do it alone, required to ask for approval, or barred from it entirely. An automatic review checks consequential actions before they run, and an activity view shows exactly what the agent did on its cloud computer. Password changes always stay with a person.
For a small business, a sensible starting setup is mandatory approval on anything that leaves the building: an email to a customer, an invoice, a payment, a social post. After a few weeks the activity view shows where the agent is reliable, and only there does it get more freedom.
Where a small business saves time
The jobs that suit an agent like this are the ones that repeat daily and need little judgement. Chasing quotes that got no reply, pulling enquiries from several channels into one list, drafting a reply to each new enquiry, turning recordings into content. The owner still approves the result in every case but skips the legwork.
Before paying, check whether the tools the business already uses are among the 4,000 apps the agent can connect to. If the invoicing software, the CRM or the calendar is missing, the agent only sees part of the picture.
At the same conference OpenAI also showed GPT-6.1 Sol, an upgraded model for professional work and computer use, and a new $500-a-month Pro tier for the heaviest users.