OpenAI introduced dots at its developer conference on September 29, 2026, putting a new always-on AI agent directly in front of Meta’s three-week-old Muse. The two products now compete for the same pitch: an assistant that keeps working after you close the app. OpenAI calls dots “remarkably capable, always-on agents built to handle everything,” and says each one runs on its own cloud computer, powered by GPT-6 Astra. That framing puts OpenAI’s agent bet on a collision course with Meta’s Muse, which launched to US users on September 8 and has already racked up download numbers that outpaced ChatGPT’s early growth curve.
What OpenAI Announced Today
Sam Altman unveiled dots during OpenAI’s developer event, positioning the product as a structural shift rather than a feature update, according to CBS News’ coverage of the launch. OpenAI describes a dot as an agent that can take ongoing responsibility for a task and keep making progress between conversations, rather than waiting for a new prompt each time. The company’s own framing, cited by PCWorld, calls dots “an extension of you,” a phrase meant to separate the product from chat-based assistants that reset context after every session.
Each dot gets its own cloud computer, a persistent environment that keeps running even when the user is offline. That detail matters because it changes the cost and reliability math for OpenAI. A cloud instance dedicated to a single user’s agent needs to stay provisioned, secured, and monitored around the clock, which is a heavier infrastructure commitment than a chatbot that spins up only when someone types a message. 9to5Google’s write-up of the rollout notes that dots reach users inside ChatGPT on web, mobile, and desktop, so the agent lives inside the same app people already use rather than a separate download.
Inside dots: Cloud Computers Powered by GPT-6 Astra
The model doing the reasoning behind every dot is GPT-6 Astra, the same system OpenAI has been positioning as its cybersecurity and agentic flagship in recent months. Pairing a persistent cloud computer with Astra’s reasoning lets a dot hold state across a long-running task, checking a calendar today and following up on a research thread tomorrow without the user re-explaining context. OpenAI’s Help Center describes a dot as able to “take on ongoing responsibility and keep making progress between conversations,” language that distinguishes it from the single-turn agents most people have used until now.
That architecture echoes, and extends, OpenAI’s earlier agentic products. The company first tested browser-based task automation with Operator in early 2025, a product that could click through websites on a user’s behalf but lacked persistent memory or a dedicated always-on environment. dots folds that browsing capability into a longer-lived agent, one that OpenAI says users can name during setup, giving it a handle in the format @yourname-agentname so it can be addressed directly inside ChatGPT, Slack, or Teams.
Meta Muse: The Agent dots Is Chasing
Meta got here first. The company introduced Muse on September 8, describing it as a personal AI agent that answers questions, browses the web, generates images, drafts documents, and connects to outside services through a system Meta calls Connectors. Muse rolled out in the United States on iOS, Android, and muse.ai, with support for AI glasses planned as a follow-up. Early traction moved fast: Muse’s download curve, covered in detail in shattered.io’s report on Muse hitting 2.8 million downloads, beat the pace ChatGPT set during its own early rollout years earlier.
Muse’s security model runs each user’s agent inside its own virtual machine in Meta’s cloud, with a permission layer called Sentinel that approves, blocks, or escalates every connected-app action and outbound network request, per CNBC’s reporting on the launch. That isolation approach became a talking point almost immediately: shattered.io’s earlier comparison of Muse’s private-VM security against three rival approaches found Meta leaning on sandboxing as its main differentiator against competitors that hadn’t yet shipped anything comparable. dots now gives OpenAI its own answer, built around a dedicated cloud computer rather than a shared or on-demand instance.
dots vs Muse: Head-to-Head Comparison
Both products chase the same job: an agent that acts on a person’s behalf without constant supervision. The specifics of how each company built that agent differ in ways that will shape which one enterprises and individual users pick first.
| Feature | OpenAI dots | Meta Muse |
|---|---|---|
| Announcement date | September 29, 2026 | September 8, 2026 |
| Underlying model | GPT-6 Astra | Not specified by Meta |
| Runtime environment | Dedicated cloud computer per agent | Per-user virtual machine with Sentinel permission layer |
| Platforms | ChatGPT web, mobile, desktop | iOS, Android, muse.ai, WhatsApp |
| App connections | 4,000+ apps via ChatGPT | Connectors to outside apps and services |
| Access tiers | Pro, Business Premium, Enterprise (beta) | Free tier plus paid plans |
| Regional availability at launch | Excludes EEA, Switzerland, UK for Pro tier | United States only |
The table above draws only on details each company has confirmed. OpenAI has not published a specific price for dots, and Meta has not disclosed which underlying model powers Muse, so both fields stay general rather than guessed. A wider view of how the two agents fared under direct testing is already up on the site: shattered.io ran Muse against ChatGPT-6 across five prompts before dots existed, and that earlier result gives a baseline for how the underlying models compared before OpenAI shipped its dedicated agent layer.
Pricing and Regional Availability
OpenAI has folded dots into its existing ChatGPT subscription structure rather than charging a separate fee, based on the tiers the company listed at launch. Access depends on which ChatGPT plan a user already holds, and the rollout isn’t uniform across regions.
| Access tier | Availability | Regional notes |
|---|---|---|
| Pro | Available at launch | Excludes European Economic Area, Switzerland, and the United Kingdom |
| Business Premium | Available across supported ChatGPT regions | No exclusions stated by OpenAI |
| Enterprise | Beta | Enabled by workspace administrators |
The EEA, Switzerland, and UK carve-out for Pro users tracks a pattern OpenAI has followed before with agentic features, where European regulatory review runs on a separate track from US and global rollouts. Enterprise access sitting in beta, gated behind administrator approval, suggests OpenAI wants IT and security teams to control the blast radius before a dot gets loose inside a company’s internal systems. That’s a far more cautious posture than the consumer-first push Meta made with Muse, which launched with a free tier open to any US adult from day one.
The 4,000-App Connector Ecosystem
OpenAI says dots can work across more than 4,000 apps through ChatGPT, a figure that reflects the connector and plugin ecosystem the company has built up since 2023. That number matters more than it might look at first glance. An agent is only as useful as the systems it can touch, and a dot that can read a calendar, file a ticket, or update a spreadsheet without the user manually copying data between tools is doing meaningfully more work than a chatbot confined to a single conversation window.
The breadth of that connector library also sets up a direct fight with Meta’s Connectors system for Muse, which links the agent to outside apps and services but hasn’t published a comparable count. Whichever company can point to more live integrations gets an easier pitch to enterprise IT buyers, who care less about raw model quality than about whether an agent can actually plug into the CRM, ticketing system, or internal wiki they already run. GPT-6 Sol and Luna, OpenAI’s other recent model releases, launched at aggressive discount pricing to undercut Claude, showing OpenAI has been willing to compete on cost as well as capability across its product line this year.
Permission Boundaries and Autonomy Controls
OpenAI built a three-tier permission structure into dots. Users can set what the agent may do entirely on its own, what requires a check-in first, and what it should never attempt regardless of instructions. That structure mirrors, in concept, the approval layer Meta built into Muse with Sentinel, where every connected-app action and outbound request gets approved, blocked, or escalated to the user.
The two companies arrived at similar answers to the same problem: an agent that can act without supervision is also an agent that can make an expensive or embarrassing mistake without supervision. Giving users a way to draw hard lines around what the agent will never do addresses the worst-case scenarios, like an agent completing a purchase it shouldn’t or sending an email that was never meant to go out. Whether either permission system holds up against real-world misuse is a separate question from whether it exists on paper, and that gap is where regulators and security researchers will likely focus first.
Security Questions Facing Always-On Agents
An agent that runs continuously on a dedicated cloud computer, with access to thousands of connected apps, is a bigger attack surface than a chatbot that only runs while a user is actively typing. Security researchers have already flagged this category of risk across the industry this year, and OpenAI’s own track record includes agent-related incidents that shaped how cautiously the company is now rolling out dots.
The gated Enterprise beta, run through workspace administrators rather than opened to every business user immediately, reads as a direct response to that risk category. An agent with standing access to a company’s connected apps needs the same kind of access review that any new employee or service account would get, and OpenAI appears to be building that review into the rollout itself rather than bolting it on afterward. For readers weighing where the broader AI agent ecosystem stands on trust and safety, the competitive landscape around GPT-6 Astra offers useful context: Anthropic has been evaluating its own response as GPT-6 Astra’s benchmark lead grows, and that competitive pressure is part of why every major lab is racing to ship an agent product this quarter.
Why the Timing Matters
Three weeks separate Muse’s launch from dots. That’s a tight window for a company the size of OpenAI to ship a competing product, which suggests dots was already deep in development before Meta’s September 8 announcement rather than a rushed response built from scratch. Still, the timing does OpenAI a favor. Muse spent its first weeks on the market absorbing press coverage, user feedback, and scrutiny from outside developers, and OpenAI gets to launch dots into a market that already understands what a personal AI agent is supposed to do.
That sequencing also puts pressure on every other lab still sitting on the sidelines. Google, Anthropic, and xAI have all shipped agentic features this year, but none has matched the always-on, persistent-environment approach both Meta and OpenAI just committed to. The next earnings call from any of those companies will likely field questions about whether they plan to match this category or cede it.
Market and Competitive Reaction
The reaction from Meta’s side of the market has already been tested once this month, when Amazon restricted Muse’s ability to operate on its retail platform. shattered.io covered the block in detail, and it’s a reminder that platform owners, not just users, get a vote on whether a third-party agent is allowed to act freely inside their systems. OpenAI faces the same dynamic. A dot that can act across 4,000+ connected apps depends on every one of those app makers continuing to allow agentic access, and any of them could pull that access the way Amazon did with Muse.
That risk cuts both ways for OpenAI and Meta. Neither company controls the full stack an always-on agent needs to function, so both are betting that the convenience an agent delivers to end users will outweigh the friction of individual platforms restricting access one at a time. So far, Meta has absorbed one high-profile platform restriction without pulling Muse back, and how OpenAI handles a similar incident, if one comes, will be an early test of how committed the company is to keeping dots broadly connected.
Historical Context: From Operator to dots
OpenAI’s agent push didn’t start this week. The company’s earlier browser-automation tool, Operator, gave users a preview of task automation in early 2025 but ran only while a user actively supervised the session, with no persistent memory once the browser tab closed. Deep Research followed as a narrower tool focused on multi-step web research rather than general task completion. dots is the first OpenAI product to combine a dedicated always-on environment with the broad app connectivity the company has been building through ChatGPT since 2023.
That progression, from supervised single-session tools to a persistent named agent with its own cloud computer, tracks the same arc every major AI lab has followed this year: ship a narrow capability, learn from how people misuse or underuse it, then widen scope once the guardrails catch up. Muse followed a similar path inside Meta’s own product line, building on years of Meta AI assistant work before arriving at a standalone agent app.
Enterprise Adoption Implications
The Enterprise beta gate is the detail worth watching most closely if you work in IT or security. Workspace administrators, not individual employees, decide whether dots gets turned on inside a company account, which puts the rollout decision in the hands of people who already manage access controls for every other piece of software the business runs. That’s a deliberate choice, and it puts OpenAI’s enterprise rollout on a more conservative footing than its consumer-facing Pro tier.
Companies evaluating dots for internal use will need to answer the same questions they’d ask of any new software with standing access to internal systems: what data can it read, what actions can it take without approval, and who audits its activity log. OpenAI’s tiered permission system gives administrators a starting point for those policies, but the burden of actually writing and enforcing them still falls on the company deploying the agent, not on OpenAI.
Five Predictions for Personal AI Agents
- Expect at least one more major lab, most likely Google or Anthropic, to announce a comparable always-on agent product within the next two to three months, closing the gap dots and Muse just opened.
- Regional rollout gaps will narrow slower than feature gaps. Expect dots to reach the EEA, Switzerland, and UK only after additional regulatory review, likely stretching into 2027.
- Platform-level restrictions like Amazon’s block on Muse will become more common, not less, as more retailers and service providers decide whether to allow third-party agents to act on their sites at all.
- Enterprise adoption will lag consumer adoption by a wide margin for both dots and Muse, as IT security teams take months to build policies around agent access before turning features on broadly.
- Pricing transparency will improve within the next two quarters. Neither OpenAI nor Meta has published a clean line-item cost for their agent products yet, and competitive pressure will likely force both toward clearer public pricing.
Frequently Asked Questions
What is OpenAI dots?
dots is an always-on AI agent OpenAI launched on September 29, 2026, built on GPT-6 Astra, with each agent running on its own dedicated cloud computer inside ChatGPT.
Is dots the same as OpenAI’s earlier Operator tool?
No. Operator was a supervised browser-automation tool that OpenAI tested in early 2025. dots adds a persistent, always-on environment and broader app connectivity that Operator did not have.
How much does OpenAI dots cost?
OpenAI has not published a separate price for dots. Access is included for Pro, Business Premium, and Enterprise ChatGPT subscribers in eligible markets, based on the tiers the company listed at launch.
Can I use dots outside the United States?
Availability depends on your ChatGPT plan. OpenAI’s Help Center specifies that the initial Pro rollout excludes the European Economic Area, Switzerland, and the United Kingdom. Business Premium is available across supported ChatGPT regions, and Enterprise access is a beta controlled by workspace administrators.
How is dots different from Meta’s Muse?
Both are always-on personal AI agents, but they differ in platform, permission architecture, and access tiers. dots runs inside ChatGPT and connects to more than 4,000 apps, while Muse runs as a standalone app on iOS, Android, muse.ai, and WhatsApp, isolated inside per-user virtual machines with a permission layer called Sentinel.
What model powers dots?
GPT-6 Astra, the same model OpenAI has positioned as its agentic and cybersecurity flagship in recent releases.
Can businesses control what a dot is allowed to do?
Yes. OpenAI lets users and administrators set boundaries covering what a dot may do independently, what requires approval first, and what it should never do, regardless of instructions given in a session.
Is Enterprise access to dots available to every business now?
Not automatically. Enterprise access is in beta and must be enabled by a company’s workspace administrators, rather than being turned on by default for every account.
