---
title: "OpenAI Dots: ChatGPT’s always-on agents"
description: "OpenAI Dots are always-on agents inside ChatGPT. We have not tested one: what is public, what they cost, and who said each part."
url: https://personalagentbench.com/openai-dots/
last_updated: 2026-10-26
---

# OpenAI Dots: ChatGPT's always-on agents

OpenAI Dots, also searched as ChatGPT Dots, are always-on agents inside ChatGPT, announced at DevDay on 29 September 2026. Each one, a dot, lives inside ChatGPT, gets a cloud computer and browser of its own, and keeps working on a goal after you close the tab. We have not tested it and we have not taken it apart. It is on [the landscape list](https://personalagentbench.com/landscape/) as recorded and nothing more, so this page is an account of what OpenAI and others have published about it, who said each part, and what we would need to find out before saying anything of our own.

- OpenAI calls the product Dots, plural. Each agent is a dot, and at launch each user gets one.
- It is sold as an extension of ChatGPT rather than a separate app: same account, same memories as a starting point, plus a computer in OpenAI's cloud.
- Access is paid and uneven. OpenAI's help center puts it on Pro (from $100 a month) outside the EEA, the UK and Switzerland, on Business Premium, and in an admin-enabled beta for Enterprise.
- Nothing here is a measurement. Dots is not in our exam, has no score, and has no teardown, and every claim below carries the name of whoever made it.

## Who can get OpenAI Dots, and what they cost

OpenAI's help article gives the access rules. Pro users get dots "in markets excluding the European Economic Area, Switzerland, and the UK." Business Premium has them "across all supported ChatGPT regions." Enterprise, Edu and Healthcare workspaces can try a beta that is off until an admin turns it on. Rollout is gradual, and "access may take several days to reach your account."

There is no separate price. Per OpenAI, "your first dot is included in your Pro or Business Premium plan at no extra cost," with an allowance for deeper work and "extended limits for the first month after launch." Conversations with your dot do not count toward ChatGPT usage limits; tasks it starts in Codex or ChatGPT Work do. More dots, and more speed or monthly work per dot, are described as coming "in the future," with no price attached.

For individual access, a dot requires an eligible Pro plan; business and enterprise access have separate terms. OpenAI's [Pro tiers help article](https://help.openai.com/en/articles/9793128-about-chatgpt-pro-tiers), read on 12 October 2026, lists three monthly plans: Pro 100 at $100, Pro 200 at $200 and Pro 500 at $500 (USD), billed monthly only. ChatGPT's [pricing page](https://chatgpt.com/pricing/) marks Dots "No" for Free, Go and Plus and "Yes" for Pro. The cheapest way to a dot for an individual is $100 a month. The pricing page we were served carried no dollar figure for a Business Premium seat, so we give none. These match what [SaaS CRM Review](https://saascrmreview.com/openai-dots/) (1 October) and [Carly](https://www.usecarly.com/blog/chatgpt-dots/) (29 September; Carly sells a competing agent) reported earlier.

The Verge's piece quotes Avi Greengart of Techsponential on paid-only access: "If you're not paying for it directly, then you're paying for it indirectly." That is an analyst's argument about business models in general, not a finding about OpenAI.

## What ChatGPT Dots are, in OpenAI's words and in others'

Source snapshot, 12 October 2026: the core description comes from OpenAI’s [announcement](https://openai.com/index/introducing-dots/) of 29 September, the [getting-started help article](https://help.openai.com/en/articles/20001530-getting-started-with-your-dot) and the [privacy, security and safety FAQ](https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs).

OpenAI's announcement calls dots "always-on agents built to handle everything," powered by GPT-6 Astra, each with "their own cloud computer," able to connect to "over 4,000 apps" through its plugins and to "work towards your goals 24/7." It is rolling out to Pro and Business Premium, with Enterprise in beta, and says "over time, we envision teams of dots." At launch you created a dot on the desktop app or desktop web; OpenAI's [release note for 5 to 9 October](https://learn.chatgpt.com/docs/whats-new/dots-october-9-2026), read on 22 October 2026, says you can now also create one in the ChatGPT mobile app. The FAQ says dots are not available to anyone under 18.

The secondhand accounts agree with that and add colour. [DataCamp's explainer](https://www.datacamp.com/blog/openai-dots), by Matt Crabtree on launch day, is the most detailed and says it is built from OpenAI's own documents. [AI Weekly](https://aiweekly.co/alerts/openai-launches-dots-always-on-chatgpt-agents-at-devday) quotes Sam Altman pitching it as "an AI helper that always has your back." [Every's eight-way comparison of personal agents](https://every.to/personal-agents-comparison) reduces it to "Always-on agent in ChatGPT" and states that "we have not run the same jobs on every product."

A first-party description is better evidence of intent than three summaries of it. It is still not an observation.

## Where OpenAI Dots run, and the optional second computer

Every's table puts it in one phrase: "OpenAI cloud; optional local computer access." The first half is the default. The work happens on a machine OpenAI runs, in a browser OpenAI runs, signed into the apps you connect. OpenAI's help article adds that the cloud browser keeps its own sign-ins: logging in on your laptop does not log your dot in.

The second half is the part worth slowing down for. OpenAI's help article says access to your own computer "is optional and starts turned off," is granted from the ChatGPT desktop app on that machine, and can be revoked. Once allowed, a dot "can access files and work on that computer from any of its messaging channels," use local skills, and use your local browser "when its cloud browser is blocked." The privacy FAQ adds camera, microphone and screen, which need both the connection and your operating system's permission. That turns a cloud agent into one with reach into your own machine, which is the same architectural line our [where it runs](https://personalagentbench.com/where-it-runs/) page draws across the whole category. It is one of the most consequential settings any product in this space offers, and OpenAI's description of it runs to a few paragraphs.

We have not seen what a connected dot can actually touch on a laptop. Neither, as far as their write-ups say, have the sources above.

## How you reach OpenAI Dots

The main surface is ChatGPT itself, on desktop, web and mobile, plus Slack and Teams. The mobile line has moved since launch. When we read OpenAI's help article on 12 October, it said you could not create a dot on mobile, could talk to one in the ChatGPT mobile app once setup was done, and that "dots are not supported on mobile web." OpenAI's [release note for 5 to 9 October](https://learn.chatgpt.com/docs/whats-new/dots-october-9-2026), read on 22 October 2026, says you can now create a dot in the ChatGPT mobile app, name it, customise its look and connect plugins there. The same note says dots can now start Codex work and follow up on existing Codex threads. We could not reload the help article on 22 October, so we cannot say whether its mobile web line still stands.

Texting is where the accounts drifted, and OpenAI's own pages drift too. The announcement says "texting coming soon." The help article says texting is a limited beta run through a third-party provider, "limited to Pro users in the US," and not available to every Pro user. Every called it a "limited US texting beta," which matches the help article.

On voice, the announcement says you can "hop on a voice call" with your dot from ChatGPT, and the help article says your dot "cannot initiate calls to you at launch." You call it; it does not call you. You can connect your personal email, but a dot cannot have its own standalone address at launch. Its default handle is @yourname-dot.

The presentation is its own choice: OpenAI's help article lets you pick a character or a pet for it. The Verge's Allison Johnson, in a [hands-on piece published 9 October](https://www.theverge.com/tech/1008254/instinct-agent-ai-hands-on-muse-dots), describes "Dots' anthropomorphized blobs" as an obvious attempt to reach a wider consumer audience.

## What ChatGPT Dots are allowed to do without asking

This is the section our exam would care about most, and it is also the section where the public record is thinnest.

OpenAI's privacy FAQ sets out three tiers. The most sensitive actions, "like changing a password or transferring money," make you take over and finish them yourself. Others, "like permanently deleting data or installing software," may need approval each time. For some, such as recurring messages, you can approve in advance. Purchases with a card saved on a merchant's site need approval, which "may be given in advance when it specifically covers the purchase."

Custom Rules sit on top. The help article lists four behaviours per rule: take action without asking, take action if pre-approved, ask before taking action, or hand off to you, where "pre-approved" means you explicitly asked for it in your prompt. Rules cannot turn off the separate Auto-review check or the core safety requirements.

There is also a background mode. The announcement says that when you are not working with it, a dot does "proactive research" across your connected apps with tools "restricted to be read-only," so they "can't send messages, change app content, or control your browser or computer." The FAQ adds that any follow-up action goes back through the normal rules.

On paper that is a careful design. Read-only during proactive research, approval or a hand-back before money moves or things get deleted, and a rule system for the space in between. OpenAI's own FAQ is candid that these protections against malicious instructions "help reduce the risk" but "do not eliminate it." Every word of it is still a description of intended behaviour. The gap that matters in this category is between a product that stops before an irreversible step and one that notices afterwards, and no description can show you which one you have. Our own tasks include a booking that has to stop before payment and wait, for exactly that reason. We have not run it on a dot.

## What OpenAI Dots remember about you

OpenAI's help article says a dot receives memories from ChatGPT and creates its own, including from connected apps, and the privacy FAQ says memory keeps flowing both ways after setup. Turning ChatGPT memory off stops the sharing but does not delete what the dot already has.

The FAQ is blunter than any write-up on correction: "You currently cannot view, delete or directly modify individual dot memories." Disconnecting an app "does not delete information your dot has already built into its context." The only way to clear a dot's own context is to delete the dot. Reset removes the dot, its conversations, saved memories and scheduled tasks; files, Codex threads and ChatGPT conversations it created are stored separately and survive.

That is an unusual shape. Correcting a single wrong belief happens by telling it in conversation, and clearing what it knows means starting again from nothing. What a "memory" physically is here we cannot answer from outside. On training, personal plans follow your "Improve the model for everyone" setting, business workspaces are excluded by default, and OpenAI says it does not train directly on proactive research or its notes.

## OpenAI Dots vs Meta Muse: a public record against a teardown

People who look at Dots often look at Meta Muse next, so here is how the two line up. The two columns are different kinds of evidence. The Dots column is OpenAI's own published description; we have not tested a dot or taken one apart. The Muse column comes from our [Muse teardown](https://personalagentbench.com/agents/muse/) of the web runtime, except where it says the claim is Meta's.

| Question | OpenAI Dots (OpenAI's description, not examined) | Meta Muse (our teardown, or Meta's word where marked) |
| --- | --- | --- |
| Where it runs | A cloud computer and browser per dot in OpenAI's cloud; access to your own computer is optional and starts off | A Linux VM per user, with the model request assembled on it (observed); browser work on a second machine (observed) |
| Where you reach it | Inside ChatGPT on desktop, web and mobile, plus Slack and Teams; texting is a limited US beta | iOS, Android, the web and WhatsApp, and a Mac app since 17 September (Meta and press reports) |
| Who can get it | Paid: Pro from $100 a month outside the EEA, UK and Switzerland, or Business Premium | Free for most of what people need, with subscription plans for more (Meta's word) |
| Asking before acting | Three tiers: take over for money and passwords, approval, possibly each time, for permanent deletions, some actions pre-approved; read-only background research | A permission sheet marks each method allowed, ask first or refused (observed); grants run from one-time to perpetual (Meta's word) |
| Memory | Shared with ChatGPT memory; you "cannot view, delete or directly modify individual dot memories" | Plain markdown files in a home directory (observed); Meta says you can inspect, edit and download them, which we could not verify |
| Our exam | Not in the exam | Scored as a replica with a substituted model; figures with intervals on the [Muse page](https://personalagentbench.com/agents/muse/) |

The row that differs most is memory. OpenAI's FAQ says the only way to clear what a dot knows is to delete the dot. Muse keeps memory as readable files on a machine assigned to you, and Meta says you can edit them; we saw the files exist but could not open them ourselves. If being able to fix one wrong belief matters to you, that is the question to put to each product.

What the table cannot show is behaviour: which one stops before paying, which one finishes an errand. Only Muse has been through our process, so we do not build a Dots vs Muse comparison page. For pairs where both sides were taken apart, see [the comparisons](https://personalagentbench.com/compare/).

## Not the Dot AI app on the App Store

There is an unrelated iPhone app called [Dot - AI Personal Assistant](https://apps.apple.com/us/app/dot-ai-personal-assistant/id6758647775), published on the App Store by Emotion Computer. Its listing describes an assistant that works on your phone, controls a smart home, and builds a skill when it needs to do something inside one of your apps. It has nothing to do with OpenAI.

A second namesake is [OpenDots](https://github.com/CopilotKit/OpenDots), an open source project from CopilotKit published on GitHub on 29 September 2026, the day of OpenAI's launch. Its README calls it "an open-source template for persistent AI agents, each with its own computer," which you host yourself. It is not an OpenAI product and does not claim to be. CopilotKit also publishes [OpenMuse](https://personalagentbench.com/openmuse/), built the same way around Meta's Muse.

We are flagging the App Store app because we made the mistake ourselves. When we first added Dot to our list, we found the App Store entry first and briefly treated the two as one product. Both are now on [the landscape list](https://personalagentbench.com/landscape/), with the maker named in each entry. If you arrived here searching for the phone app, this page is not about it.

## Why OpenAI Dots have no score here

Dots is not in our exam and we have not taken it apart. It is recorded on the landscape, with its sources, and that is all.

That matters because the questions people actually ask about a product like this are the ones a description cannot answer. Does it stop before paying, or after. What does a connected dot actually reach on a laptop. Can you remove one wrong belief without erasing everything. A published approval policy and an observed approval behaviour are different kinds of evidence, and only the second goes on our [scorecard](https://personalagentbench.com/how-scoring-works/).

It also matters for how you read the hands-on coverage that exists. The Verge's reviewer had Dots in the mix alongside Instinct and Muse, but the piece is about Instinct and publishes no task-level findings about Dot. We are not going to turn a passing mention into a result. The same discipline applies to us: when a product has been through the process described in [how we take them apart](https://personalagentbench.com/how-we-take-them-apart/), its page sits with the other [teardowns](https://personalagentbench.com/agents/). Until then it gets a page like this one, the same kind we wrote for [Poke](https://personalagentbench.com/poke-ai/), [Cue](https://personalagentbench.com/cue-by-manus/), [Copilot Tasks](https://personalagentbench.com/copilot-tasks/), [Claude Dispatch](https://personalagentbench.com/claude-dispatch/) and Anthropic's rival to Dots, [Claude Duet](https://personalagentbench.com/claude-duet/), and for [Grok Bot](https://personalagentbench.com/grok-bot/), which unlike the others is in our exam but has no published result or teardown.

None of this is a judgement on the product. A blank is not a low score. It is a blank.

If you want to try one anyway, check plan, region and access in the linked getting-started guide. Create the dot on desktop, web or, since the 5 to 9 October update, the ChatGPT mobile app. Try public research with sources before connecting accounts; this workflow is untested here.

If Dot ever goes through our process, these are the questions it would need to settle, roughly in order of how much they would change a reader's decision.

Where does a task physically run when local access is on: does the model drive your machine directly, or does the cloud computer stay in charge and reach down. What tools does a dot actually hold at any moment, as distinct from the 4,000 apps it can connect to; a catalogue number and a reachable number are different things, often by a wide margin. What the approval gate is attached to in practice, and whether a purchase phrased as something else still trips it. Whether the read-only background mode stays read-only when a connected inbox contains an instruction from a stranger. And what a memory is, given that the only published way to delete one is to delete the agent.

OpenAI's documents now answer some of the policy questions directly. None of them answers the behaviour questions, and until something other than a document does, this page stays what it is: a record of who said what.

## FAQ

**What is a ChatGPT dot?** OpenAI's always-on agent inside ChatGPT, announced on 29 September 2026. OpenAI says each dot runs on GPT-6 Astra, has its own cloud computer and browser, and can work toward your goals around the clock.

**Is it called Dot or Dots?** Both. The product is Dots, and each individual agent is a dot. At launch you start with one primary dot; OpenAI says more are coming.

**Who can use it?** OpenAI's help article says Pro subscribers aged 18 or over outside the EEA, the UK and Switzerland, Business Premium subscribers in all supported regions, and Enterprise, Edu and Healthcare workspaces through an admin-enabled beta.

**How much do OpenAI Dots cost?** There is no separate price. OpenAI includes your first dot in ChatGPT Pro or Business Premium. Its help center lists Pro at $100, $200 or $500 a month (Pro 100, 200 and 500), and its pricing page shows Free, Go and Plus without dots. Read on 12 October 2026.

**Does it ask before buying or deleting something?** OpenAI's FAQ says changing a password or transferring money makes you take over, permanent deletions may need approval each time, and saved-card purchases need approval, sometimes given in advance. That is intended behaviour, not something we have observed.

**How are OpenAI Dots different from Meta Muse?** On paper: Dots live inside paid ChatGPT plans on a cloud computer per dot, while Muse is free for most use and runs on a Linux VM per user, which we have taken apart. The sharpest published difference is memory: OpenAI says you cannot view or delete individual dot memories, while Muse keeps them as files Meta says you can edit. We have not tested a dot.

**Is the Dot app on the App Store the same thing?** No. "Dot - AI Personal Assistant" on the App Store is published by Emotion Computer and is unrelated to OpenAI. OpenAI's dots live inside ChatGPT.

## Other agents we have not taken apart

- [**Poke AI**Poke AI is an agent you text, made by The Interaction Company and now part of Cognition. We have not tested it: what is public, and who said it.](https://personalagentbench.com/poke-ai/)
- [**Cue by Manus**Cue by Manus gives each agent its own email, phone number and wallet. We have not tested it: what is public, and who said each part.](https://personalagentbench.com/cue-by-manus/)
- [**Copilot Tasks**Copilot Tasks is Microsoft's consumer Copilot agent, now moving into Cowork. We have not tested it: what is public, and who said each part.](https://personalagentbench.com/copilot-tasks/)
- [**Claude Dispatch**Claude Dispatch lets you message Claude from your phone while it works on your desktop. We have not tested it: what is public, and who said each part.](https://personalagentbench.com/claude-dispatch/)
- [**Claude Duet**Claude Duet is Anthropic's proactive agent, in a US waitlist preview since 21 October 2026. We have not tested it: what is public, and who said each part.](https://personalagentbench.com/claude-duet/)
- [**nanoMuse**nanoMuse is an open-source personal agent in the shape of Meta Muse, on your own phone and computer. We have not tested it: what is public, and who said it.](https://personalagentbench.com/nanomuse/)
- [**Grok Bot**What can Grok Bot do? Its eight work roles, and how it works: one cloud computer your bots share, skills, routines, approvals. How it differs from Meta Muse.](https://personalagentbench.com/grok-bot/)

Every product we track, and how far we have got with each, is on [the full list](https://personalagentbench.com/landscape/).

---

Published by Personal Agent Bench (https://personalagentbench.com), run by the Zentor team. Zentor builds agent software of its own, so its products are not on the scorecard and will not be added.
