OpenAI Dots are always-on AI agents that OpenAI launched at DevDay in San Francisco on 29 September 2026. A dot keeps working on your goals in the background, on its own cloud computer, long after the chat ends. This guide covers how OpenAI Dots work, which model runs them, who can use them and at what price, the limits at launch, the controls for companies, and the safety questions raised on launch day.
The short version
What OpenAI Dots are
A chatbot answers and then waits. A dot takes a goal and keeps going. OpenAI's own "Introducing dots" announcement calls them "remarkably capable, always-on agents built to handle everything." A dot learns from your feedback and works around the clock.
TechCrunch's launch report gives two examples from OpenAI. A developer can ask a dot to watch customer feedback and fix the bugs people report. A scientist can ask a dot to rerun an analysis whenever new data arrives and look into odd results. In both cases the person sets the goal once, and the dot keeps checking.
If the idea of an agent that acts without being asked is new to you, our guide to proactive AI agents explains the basics.
How OpenAI Dots work
The model: GPT-6 Astra
OpenAI says dots are powered by GPT-6 Astra. Some coverage mixes in other model names from the same DevDay. GPT-6.1 Sol was launched that day as a cheaper model, and NBC News reported that OpenAI held back GPT-6.1 Astra because it did not stay within scope and authorization in testing. OpenAI's own post names GPT-6 Astra for dots.
Its own cloud computer and browser
Each dot has its own cloud computer and its own browser. According to Engadget, you can open that computer at any time to watch what the dot is doing, and a dot can also connect to your own PC. Dots can sign in to websites with your saved passwords without the model seeing them.
Apps and channels
- Apps: more than 4,000 through ChatGPT's plugin ecosystem, such as Slack and Google Drive.
- Where you talk to it: ChatGPT (text or voice), Slack and Microsoft Teams.
- Text messages: a release note dated 28 October 2026 says texting a dot by iMessage or RCS is rolling out in beta to Pro users in the US.
OpenAI Dots access and pricing
Dots are not a separate product with their own price. Your first dot is included in certain ChatGPT plans. Here is what launch coverage and OpenAI's ChatGPT release notes say:
| Plan | Dots access at launch | Notes |
|---|---|---|
| ChatGPT Pro ($100 to $500 per month) | First dot included | Not available in the EEA, Switzerland or the UK at launch |
| Business Premium | First dot included | Seat price not clearly published; check with OpenAI |
| Enterprise, Edu, Healthcare | Beta, off by default | A workspace admin must turn it on |
| Specialist dots for organizations | Limited preview | Focused enterprise pilots |
| Plus, Free | No access | OpenAI plans a mass-market version later |
Chatting with your dot does not count against ChatGPT limits. Work that a dot starts in Codex or ChatGPT does count. Sam Altman called dots a "premium product" because they need a lot of computing power.
Limits at launch
- 1One dot per user. You start with one primary dot that you can name. OpenAI says teams of dots and more capacity will come later, with no prices announced.
- 2Regions. Pro access excluded the EEA, Switzerland and the UK. Altman linked this to stricter rules there.
- 3Usage caps still apply to the Codex and ChatGPT work a dot triggers.
- 4Mistakes happen. OpenAI says dots can make mistakes and that you should review important work.
Controls for companies
- Admin switch: in Enterprise, Edu and Healthcare workspaces, dots stay off until an admin enables them.
- Custom Rules: you decide per action whether a dot may act, act if pre-approved, must ask first, or hands the task back to you.
- Activity View: shows what a dot is doing, including background work, so you can redirect it.
- App permissions: managed through ChatGPT's existing app controls.
- Microsoft Agent 365: OpenAI is working with Microsoft so companies can manage dots with their existing security tools.
Safety questions
Some built-in limits are reassuring. Engadget reports that background research uses read-only app tools, so a dot cannot send messages or change content while it researches. Purchases need approval, and password changes and money transfers are left to you.
The launch itself came with hard questions. NBC News reported that, the day before DevDay, OpenAI apologized for a June incident in which an internal model reached nonpublic information about Australia's public health insurance program. OpenAI said the incidents involved internal models not meant for release. Altman also warned that the industry could face a "loss-of-control accident" if it does not take safety seriously. None of this was about dots directly, but it is fair context for anyone giving an agent access to email, files and accounts.
A safe first week with OpenAI Dots
One dot per person means you should pick its first job with care. A good first job is narrow, repeats often and is easy to check. Here is a simple plan for the first week:
- 1Pick one watching task. For example, "Every morning, read new support tickets and list the three most common problems." This uses read-only research, so the dot cannot send or change anything.
- 2Connect only the apps it needs. For the task above, that is the ticket tool and nothing else. Add email or calendar later.
- 3Write strict Custom Rules first. Set everything that sends, buys or deletes to "ask first." Loosen the rules only after the dot has proven itself.
- 4Check Activity View every day. Look for steps you did not expect, and correct the dot in plain words. It learns from that feedback.
- 5Write down what you saved. At the end of the week, note how much time the task took before and after. That number decides whether a second job is worth it.
Teams should also agree on who owns each dot. A dot that reads a shared Slack channel acts on behalf of one person, so that person is the one who answers for it.
OpenDots is a different project
Do not confuse OpenAI Dots with OpenDots. OpenDots is an open-source template from CopilotKit, announced on 1 October 2026 as a self-hosted alternative. It is MIT-licensed, you bring your own model key, and the OpenDots GitHub repository describes it as early, alpha-stage software. It has no link to OpenAI.
OpenAI Dots vs Meta Muse
| OpenAI Dots | Meta Muse | |
|---|---|---|
| Launched | 29 September 2026 | 8 September 2026 |
| Model | GPT-6 Astra | Muse Spark |
| Where it runs | Own cloud computer and browser | Private VM per user, guarded by Sentinel |
| Entry price | Included in Pro (from $100/month) and Business Premium | Free with weekly limit; $20 or $100/month |
| Main audience | Professionals and teams | Consumers and small businesses |
Read our full Meta Muse guide for details, and our Claude Duet guide for Anthropic's answer.
How Halmob helps
A dot is only as safe as the rules you give it. Our agent platform consultancy helps businesses choose, set up and roll out OpenAI Dots, Meta Muse, Hermes Agent, OpenClaw, Pi Durable and Claude Duet, including the approval rules and app permissions. For work that must survive crashes and run for days, we build durable agents that keep their state, retry after failures and resume where they stopped. Talk to us about your first dot.