OpenAI dots vs Meta Muse: whose computer, whose keys, whose approval

OpenAI's dots page, published on 29 September 2026, calls them "always-on agents" and says they are powered by GPT-6 Astra. Meta's Muse page, read on 24 October 2026, calls Muse a personal agent that "gets to know you and your goals, working across all aspects of life". Strip the tone and the two products make the same offer: something that keeps working after your laptop closes, remembers what you told it, and acts on your behalf inside the apps you already use. What separates them is not that offer. It is the plumbing underneath it, and the plumbing is where the pages get uneven — whose computer the agent runs on, where you are allowed to reach it, who ends up holding your passwords and your card, what has to happen before it acts, what each company says about your data, and what either company is willing to have measured. On several of those questions the honest answer is that nothing has been published, and this comparison marks those gaps rather than papering over them.

One model name, and a page that declines to give one

OpenAI names what is doing the reasoning. Dots are "powered by GPT-6 Astra", according to the page published on 29 September 2026. Meta's Muse page, read on 24 October 2026, describes what the agent does and what it runs on but does not name the model that powers it. That is a real asymmetry rather than a quirk of copywriting, because a model is the one component of an agent stack a buyer can independently look up, price and compare. Everything else on either page is a claim about behaviour; only one of the two companies puts a nameable object on the table.

The second asymmetry is the shape of the offer. OpenAI says dots are rolling out "across Pro, Business Premium, and Enterprise plans in eligible markets", and shows a preview of specialist dots with their own identity for access management, IT-provisioned hardware and deep integrations with a company's systems of record, naming a Microsoft Agent 365 integration. Meta's page is built around individual Muses — creator stories introduce four by name, Pumpkin, Lux, AunTEY and Mango — plus a Muse for Small Business described as "a collection of new skills and connectors". Neither page publishes a price. For both products, a price for the agent itself is not published, so any cost comparison a reader might want is unavailable today.

What runs the agent

This is the clearest published difference between the two products, and it is a difference about custody of a machine rather than about capability.

OpenAI dots. The dots page says each dot works on its own cloud computer, and that your computer and its contents stay separate unless you choose to connect it. The sentence does three things: it puts the work somewhere other than your laptop, it promises separation by default, and it makes connection to your own machine an explicit choice. It is also what lets the page describe a dot as working towards your goals 24/7 without asking your computer to stay awake.

Meta Muse. Muse comes with "a secure computer with a browser shared by you and your agent". The phrase "shared by you and your agent" is doing the work here. This is not framed as a machine that belongs to the agent and sits apart from you; it is a shared surface, and the browser is part of the description rather than an add-on.

Where the pages stop. Meta does not publish where that shared computer physically sits — on your Mac, on Meta's servers, or somewhere else — and does not publish what happens to its contents when the agent is idle. OpenAI does not publish the specifications of the cloud computer a dot receives, and it does not publish a per-dot allowance. Those details are not published on either side, and they matter more than they sound: they decide how much work an agent can genuinely take on before it starts handing tasks back to you.

How credentials are handled

The two pages protect your login from different parties, and that is the most useful distinction in either security section.

OpenAI dots. For signing into supported websites, dots "can use saved passwords without exposing them to the model". The protected party is the model: the dot authenticates, and the language model driving it never sees the password. What OpenAI does not publish is what happens when a site has no saved password, whether a dot can create and store a new credential, or how saved passwords are kept and audited on the dot's own computer. None of that is published. Nor is any card handling: what a dot may do with a payment method is not published on the dots page.

Meta Muse. Logins are kept in "a secure credential store your agent can't read", with 1Password integration "coming soon". The protected party here is the agent. That is a different claim from OpenAI's: it constrains what the agent itself can read, not only what the model can see. Meta also handles the card. At checkout a one-time card number is generated "so your real card is never seen by the merchant or your agent", and eligible purchases are covered by Link's purchase protections.

What neither page says. Neither vendor publishes what happens when a credential is revoked in the middle of a task, or whether either agent can be told to forget a login it has already used. Those are not published either, and they are the questions worth asking before you hand over an account rather than after.

Where you reach it

OpenAI dots. A dot is reachable through ChatGPT, Slack or Teams, and the page invites you to "hop on a voice call" when you want to talk a problem through. The design assumption is workplace-shaped: the agent arrives in tools a team already has open rather than asking anyone to install another window.

Meta Muse. Muse runs as the Muse app on mobile and Mac and directly inside WhatsApp, and the page's download prompt is for Mac. The design assumption is conversation-shaped and personal: the same kind of chat you already have with people, with the agent inside it.

The gaps are mirror images. OpenAI does not publish whether dots has a standalone app of its own; Meta does not publish whether Muse can be reached from Slack or Teams. Both are not published. What each company does publish is a list of surfaces, and a list of surfaces is a promise about where the agent will be, not a statement that the list is complete. Treat both as partial by construction.

What approval looks like

OpenAI dots. Dots "start with built-in rules for when to act independently and when to ask for approval". Custom Rules let you allow an action, require approval for it, or block it. Before an action that could affect your accounts or share information, an auto-review checks it against your instructions, your Custom Rules and the company's safety requirements, and sorts the work into what can proceed, what needs approval and what you must do yourself. Certain sensitive tasks, "such as changing a password", always stay with you. Progress, including background work, appears in an Activity View, and monitoring can pause or stop a dot's work. The page is also unusually blunt about the limit: "Dots can still make mistakes, so always review consequential work."

Meta Muse. "Review and approve actions from Muse before they happen, like sending emails and making purchases. See a complete audit trail" of everything the agent has done and what it is planning to do. The published picture is a human gate placed before the action and a record kept after it.

Where the judgement sits. OpenAI describes a rules engine plus an automated reviewer, with a person pulled in for a set of sensitive tasks it names by example. Meta describes review by the person, with an audit trail as the evidence. Neither publishes the complete list of actions that always require a human. OpenAI gives changing a password as its example; Meta gives emails and purchases. The full always-ask or refusal list is not published on either side, and that list is exactly what you would want before deciding how much rope to give an agent.

What each vendor says about training data

OpenAI dots. The dots page makes two written commitments. Content from ChatGPT Business, Enterprise and Edu workspaces is not used "to improve our models by default", and OpenAI does not train "directly on proactive research or your dot's notes to itself". The second commitment covers two categories of agent-generated material that the page names explicitly, which is a narrower promise than "we never train on your data" and a more specific one.

Meta Muse. The Muse page's data statement is about advertising: "your conversations are not shared with Meta's ad systems." That is a statement about ad targeting, not about model training, and the two are not interchangeable. Whether conversations with Muse are used to train models is not published on the page read on 24 October 2026.

Read side by side, the difference is one of subject matter. OpenAI addresses training and names which workspace data and which agent-generated material are excluded. Meta addresses advertising and says conversations stay out of the ad system. Neither sentence answers the other company's question, so a buyer who cares about training data gets an answer from one page and a gap from the other. The gap is not a denial, and it should not be read as one.

What is measured

Neither company publishes an independent benchmark for its agent, and neither publishes a price. Those are the first two numbers a buyer would reach for, and both are not published.

What each does publish is instrumentation. OpenAI's Activity View shows progress including background work, and the auto-review decides what proceeds, so you get a running view of what a dot is doing. Meta publishes a "complete audit trail" of what the agent has done and what it is planning to do. Both are records of activity. Neither is a measurement of how often the agent gets a task right.

The remaining blanks follow the same pattern. OpenAI does not publish a per-dot allowance or a dated list of which markets are eligible today. Meta does not publish which model powers Muse or a country list. For OpenAI, the plan tier is the only cost signal on the page, and it is not a price for dots. For Meta, there is no cost signal at all. The practical consequence is that these two products cannot be compared on capability or cost from the published record. They can be compared on design: where the computer is, who can read the credential, what gate stands before an action, and what each company commits to in writing about training.

What the gaps mean if you are choosing

The two products are uneven in opposite directions. OpenAI publishes the model name, the machine model, the rule system and the training exclusions, and leaves out the price and the market list. Meta publishes the credential design, the checkout mechanism and the audit trail, and leaves out the model and the price. If your decision turns on knowing what is doing the reasoning, only one page answers. If it turns on what happens to your card number at checkout, again only one page answers. Neither answers the question that usually decides a purchase, which is what the thing costs.

There is also a difference in what each company treats as the risk to be solved. The dots page is written around what the agent might do: read-only background research, rules for acting independently, an automated reviewer, monitoring that can stop the work. The Muse page is written around what the agent might see: a credential store it cannot read, a card number the merchant never sees, conversations kept out of the ad system. Those are two different threat models, and they produce two different products.

• Ask where the agent's computer physically lives before you connect anything to it.

• Ask whether the credential store is one the agent can read, not only one the model can see.

• Ask for the complete list of actions that always require your approval, not the examples on the landing page.

When the answer is silence, record it as not published and price the uncertainty into the decision. A vendor that has not published a limit has not promised one, and an agent is only as safe as the smallest thing it is allowed to do without asking.

If what you need is the model layer rather than either packaged agent, OrcaRouter serves 200+ models through one OpenAI-compatible endpoint and one API key, with automatic failover and 0% markup — provider list price passed through. Claude Opus 5.5 is priced at $4.00 per million input tokens and $20.00 per million output tokens, with a 1M-token context window and cache reads at $0.20 per million. We host models, not apps: dots and Muse are not available through us.