--- title: "Same Skeleton, Different Bets: Claude Duet vs Meta Muse vs OpenAI dots" author: "Samod Alex" published: "2026-10-25T13:02:59.671Z" updated: "2026-10-25T13:03:18.760Z" url: "https://zyvop.com/same-skeleton-different-bets-claude-duet-vs-meta-muse-vs-openai-dots-jjjz8" canonical: "https://zyvop.com/same-skeleton-different-bets-claude-duet-vs-meta-muse-vs-openai-dots-jjjz8" tags: ["meta-muse","Anthropic","agent-security","claude-duet","comparison","AI agents","openai-dots"] --- # Same Skeleton, Different Bets: Claude Duet vs Meta Muse vs OpenAI dots All three run on their own cloud computer and ask before they pay. What separates them is who checks the agent, what it costs in usage, and who can get in. ![Cover image]() Three companies shipped an always-on personal agent in six weeks. **Meta launched Muse on September 8**, **OpenAI launched dots on September 29**, and **Anthropic opened Claude Duet on October 21**. Each gets its own cloud computer, a browser, a large set of connectors, and a promise to ask before it spends your money. On a feature list they look interchangeable. This piece compares them where they diverge: access, cost, who checks the agent, what you can watch, and what happens to your data. Nothing was run hands-on. Duet is four days old and waitlisted, so the comparison rests on vendor documents, reporting, and one early outside review. > **Quick take:** Start with **Claude Duet if you pay for Claude in the US**, but keep it read-only because it **cannot show you its browser**. Choose **OpenAI dots if you need Slack, Teams, or a team route**, from **$100 a month**. Keep **Meta Muse to personal errands for now**: it is free and well documented, but it paused chat sharing on **October 20** after a data exposure. * * * ## Three launches, one skeleton Each agent works on a remote machine, asks before it sends, pays, signs, or deletes, and says the model never directly sees your passwords. The table shows where they differ. | | Claude Duet | Meta Muse | OpenAI dots | | --- | --- | --- | --- | | Model | Claude Opus 5.5 | Muse Spark 1.3 | GPT-6 Astra | | Access | US Pro and Max, by waitlist | US and Canada, free tier with card on file | ChatGPT Pro outside the EEA, Switzerland, and UK; Business Premium; Enterprise beta | | Cheapest route | Pro at $20, included in the preview | Free; Power $20; Maximum $100 | Pro 100 at $100, first dot included | | Where you talk to it | Own iOS and Android app, claude.ai | muse.ai, iOS, Android, WhatsApp | ChatGPT desktop, web, mobile; Slack; Teams | | Watch it work | No browser view or takeover yet | See the browser, take over any time | Open the dot's computer; Activity View on desktop | | Team path | Not yet, planned for later | Small-business version launched separately | Business Premium, Enterprise beta | _Sources: Anthropic's_ [_launch post_]() _and_ [_help center article_]()_, OpenAI's_ [_launch post_]()_, Meta's_ [_security post_]()_,_ [_TechCrunch's launch coverage_]()_, and_ [_Fox Business on the small-business version_]()_._ * * * ## Who gets in, and what it costs **Claude Duet** is a research preview for **Pro and Max subscribers with a US account, by waitlist**, and Team and Enterprise come [later](). It is included at no extra charge, so the price is a shared usage allowance. Its check-ins count against the same five-hour and weekly limits as regular Claude, which is why Anthropic checks accounts only **every few hours**. **OpenAI's dots** start at **$100 a month** with ChatGPT Pro, per [OpenAI's Pro tiers page](), and Pro excludes the EEA, Switzerland, and the UK, per [its setup article](). Business Premium covers all supported regions at [$125 per user per month](). Plans carry **extended limits for the first month after launch**, per [OpenAI](), so costs may shift around the end of this month. **Meta's Muse** has a free tier that needs a card on file, plus [Power at $20 and Maximum at $100 a month](). It is in the US and, per [TechCrunch](), Canada. Sensor Tower estimates [10.9 million downloads in 40 days](), which measures installs, not finished tasks. No comparable figures exist for dots or Duet. * * * ## Proactivity and early evidence Duet's pitch is judgment: a morning brief, watched accounts, and a message only when something needs you. Anthropic admits it can [misread a thread or misjudge urgency](). Every's Dan Shipper called it ["the first personal agent I'd call a pro tool"]() after two weeks of early access, but listed gaps: it cannot direct his Claude sessions, its email drafts trail his dot's, and scheduled runs cannot reach custom connectors. Anthropic gave Every early access and had no input. OpenAI sets the boundary for background work in code. A dot's [proactive research]() is read-only, so it cannot message anyone, change content in connected apps, or control a browser. Anthropic's material does not say whether Duet's check-ins are restricted the same way. Muse surfaces a feed of unprompted suggestions. * * * ## Who checks the agent Each vendor says the agent asks before acting. The useful question is what decides, and whether the agent can talk that decider out of its answer. ```mermaid flowchart TB subgraph MUSE["Meta Muse"] direction LR M1["Agent in runtime cell"] -->|"proposes action"| M2["Sentinel: separate host-side agent"] M2 -->|"allow or deny"| M3["Connector call or network egress"] M2 -->|"ask"| M4["Approval dialog in the app"] end subgraph DOTS["OpenAI dots"] direction LR D1["Dot on its own cloud computer"] -->|"plans a step"| D2["Auto-review: separate safety system"] D2 -->|"allowed"| D3["Step runs"] D2 -->|"blocked"| D4["Reason returns to the dot"] end subgraph DUET["Claude Duet"] direction LR C1["Duet on its own cloud computer"] -->|"proposes action"| C2["Auto mode classifier and approval rules"] C2 -->|"read and draft"| C3["Runs on its own"] C2 -->|"send, pay, sign, delete"| C4["You approve the exact action"] end ``` Meta's design is the most explicit. Sentinel is a [separate host-side agent and the sole permission authority]() for connector actions and network egress: "Muse proposes actions, but only Sentinel can grant permission to perform action." Kernel-level tracking marks processes that have read user data, so tainted requests fall back to approval. OpenAI's Auto-review is [a separate safety system]() that checks each planned step against your instructions and rules. A blocked step goes back to the dot with a reason, and the controls sit outside the environments dots can change. Per Anthropic's launch post, Duet's actions are approved by the same [auto mode classifier]() as Claude Code, plus a new classifier for sensitive information. Anthropic has not published Duet's version, so the details here are assumed to carry over from Claude Code, where the classifier never sees tool results and so cannot be argued into a bad call by hostile text. For coding, Anthropic published error rates: **0.4%** of benign actions blocked on 10,000 real tool calls, and **17%** of 52 real overreaching actions let through, which the post calls "the honest number." No Duet-specific rate is public, and the promised Duet safety write-up had not appeared by October 25. Read side by side, Meta leans on deterministic boundaries, OpenAI mixes code limits with a separate review, and Anthropic leans on classifiers and approvals. That is a reading of vendor documents, not a test. You can watch two of the three. Anthropic's help center says you [cannot watch Duet work in its browser or take over](). OpenAI lets you open a dot's computer at any time, and Meta shows Muse's browser and lets you take over. * * * ## Credentials, money, and limits Passwords stay out of the model's context in every design. Duet's handling [follows Claude Managed Agents](), OpenAI pauses the model during secure login, and Meta's authd service hands the agent only surrogate tokens. Meta alone describes filtering one-time codes and reset links out of the email connector. The other two documents describe no equivalent, which is not proof that none exists. **Duet and Muse both pay through Stripe's Link**, which issues a one-time virtual card, and Meta asks approval on every payment. Dots buy with saved merchant cards, with approval. Anthropic lists trading securities and moving money between accounts as off-limits even with approval, and OpenAI hands password changes and account-to-account transfers back to you. Both Anthropic and OpenAI also gate sensitive data: Duet's classifier rejects sharing card numbers and government ID, and dots need a named recipient for health data. * * * ## Data and the trust record Duet uses your Claude account settings, and disconnecting an app [keeps memory of earlier information](). A dot keeps its own context, which you can reset, and deleting the dot removes its memories. OpenAI does not train on Business, Enterprise, or Edu content by default, and the material reviewed describes no Duet-specific training policy. Meta stores everything in your VM, where you can inspect and download it, and uses sanitized trajectories for training by default, with an opt-out. Staff access is limited by policy, not prevented, until a Confidential VM arrives later this year. Muse then had a public failure. Claroc researchers found about **6,800 shared-chat links** listed in public DNS records, because Muse put each conversation's title in the share link's hostname. A minority of the few hundred pages they opened held a loan application, a bank fraud alert, and an internal revenue review. Claroc sells tools that detect public AI share links, so it has a stake, and it found no evidence of misuse. Claroc's CTO says he reported the issue to Meta's bug bounty on October 8 and heard nothing. Meta [paused the feature]() once Ars Technica asked about it, on October 20. Sampled pages stopped opening, but the titles remain in passive DNS history, per [PPC Land](). The flaw sat in the sharing feature, outside the VM and Sentinel described above, and Meta's statement did not confirm the October 8 report. Within two weeks of launch, Amazon [blocked Muse]() from its store, saying it had not been told and had not authorized access. * * * ## What nobody has measured No independent head-to-head test exists. [VentureBeat]() noted that Anthropic published no task-level cost, productivity, or benchmark results for Duet, and the numbers that do exist measure different things: installs, classifier error on coding sessions, and design descriptions. Anecdotes cut both ways. Shipper said his own dot lost a tax task without alerting him, which is one user's experience and points to a failure mode to test, not a rate. Early coverage also contained errors, such as an aggregator [swapping Muse and dots between OpenAI and Meta](). * * * ## Which one to try first **The quick take's ordering holds after the evidence, with one caveat:** it is a judgment from documents and early reporting, not testing, and it could change within weeks. ```mermaid flowchart TD A["What do you need?"] --> B{"Slack, Teams, or a team workspace?"} B -->|"yes"| D["OpenAI dots: from $100 a month"] B -->|"no"| C{"Paying for Claude in the US?"} C -->|"yes"| E["Claude Duet waitlist: no extra charge, read-only first"] C -->|"no"| G{"Personal errands only, free to try?"} G -->|"yes"| H["Meta Muse: US and Canada, card required"] G -->|"no"| D ``` - **Duet:** no extra cost, hard off-limits, ask-first gates, and a positive early review. It is weak on the invisible browser, the waitlist, and a classifier measured only on coding. - **dots:** the most complete controls for work, with a computer you can open, a read-only background mode, and a Slack, Teams, and Business route. It costs $100 and up. - **Muse:** free with a detailed security design, but it had a sharing exposure in six weeks, 12 days between report and pause, and an Amazon dispute. Revisit it for business data after Meta's review. Whichever you pick, start read-only: morning briefs, inbox triage, and drafts you review. Add one write permission at a time, keep payments on a one-time card, and read each approval prompt, as Anthropic's [help page]() advises. Details reflect launch documents and reporting as of October 25, 2026, and the Duet preview, the dots limits, and Meta's review are all still moving.