OpenAI's Dots Are Always-On Agents With Their Own Cloud Computer. Here's What They Can Do, What They Hand Back to You, and What Isn't Clear Yet

OpenAI's own announcement pages returned an access error when we tried to load them, so the launch details below come from The Verge, MacRumors, and OpenTools, plus a DEV Community write-up that cites OpenAI's Help Center. Where a detail comes only from one of those, we say so.

What a Dot is

A Dot is meant to be an ongoing assistant rather than a chat session. According to The Verge, each Dot is powered by GPT-6 Astra, has its own cloud computer with a web browser, and can reach more than 4,000 supported apps. You message it in a text-message-style interface, or hop on a voice call from ChatGPT on web, desktop, or mobile. It can also connect to Slack and Microsoft Teams and carry context across them. Once you give it a task it messages you with updates and questions, and you can open its cloud computer to watch or step in. Texting is listed as coming soon.

Who gets it, and what counts against your limits

• Rolling out from September 29 to ChatGPT Pro, Business Premium, and Enterprise, in select markets. Per the DEV Community write-up of the Help Center, Pro excludes the EEA, Switzerland, and the UK, and Enterprise, Edu, and Healthcare workspaces get an admin-enabled beta.
• One Dot per eligible user for now, at no extra cost. OpenAI says you will later be able to add more Dots and scale each one's speed or monthly work.
• Conversations with a Dot don't count toward ChatGPT usage limits. Tasks a Dot starts in Codex or ChatGPT Work still count, per OpenTools.
• Setup happens on desktop (the ChatGPT desktop app, or chatgpt.com/dots on desktop); after that you can use the mobile app.
• OpenAI's CFO Sarah Friar told CNBC the vision is to bring Dots to the whole consumer base, according to The Verge.

What it will and won't do on its own

OpenAI says Dots use "built-in rules for when to act independently", and you can add your own. The same Help Center material says local computer access is optional and starts off, and that the protections against prompt injection "help reduce the risk" but "do not eliminate it".

What the launch doesn't settle

• Real-world reliability. OpenAI's examples include a developer asking a Dot to build and test updates from customer feedback, and OpenAI says its engineers use Dots to fix dozens of bugs a day (per OfficeChai's liveblog of the keynote). These are OpenAI's own descriptions, not independent completion rates.
• Memory control. According to the DEV Community summary, you can't view, edit, or delete individual memories; you can pause a Dot or reset it, which deletes conversations, saved memories, and scheduled tasks.
• The limit on deeper work. OpenTools notes OpenAI hasn't published a simple entitlement for how much work a Dot can do; limits are extended during the first launch month.
• Specialist Dots. OpenAI is testing role-based Dots for organisations (it named accounting, email marketing, legal analysis, and ticket resolution on stage), which remain focused pilots.
• Safety in the wild. The Verge notes that rival Meta's Muse has drawn safety concerns, including one user who said it gave out their address. OpenAI points to its review and rules system as its answer.

What this means if you're choosing an agent

If you already pay for Pro or Business Premium, the first Dot is free to try, and the useful test is narrow: give it one bounded responsibility, connect the fewest apps it needs, and read its first completed task in the activity view before widening access. Teams on Enterprise need an admin to turn it on. If you want agents on your own infrastructure instead, the control-plane approach in OpenClaw Enterprise, also announced this week, is aimed at that case.