Claude Duet vs. OpenAI Dots vs. Meta Muse: Always-On Agents Compared
Three labs now ship agents that work while you are away. Here is how Duet, Dots and Muse for Small Business differ on access, approvals and cost.
Jump to 9 sections
Claude Duet, OpenAI Dots and Meta Muse for Small Business are all agents that keep working while you are away. Duet checks your apps every few hours and needs approval for sends and payments. Dots runs on its own cloud computer with GPT-6 Astra. Muse is mostly free and built around small-business tools like Shopify and QuickBooks.
In the span of about three weeks, three major labs shipped the same idea: an agent that does not wait for a prompt. The pitch is the same, and the fine print is not. Model Drop covers the apps and developer tools built on new models, and this comparison is for owners and teams deciding which agent deserves access to their email, books and customer data.
What Do These Three Agents Have in Common?
They share a basic design: connect your apps, let the agent run in the background and approve what matters. Each one gets its own cloud environment, each one reaches into business software and each one asks for sign-off before it does anything consequential.
The common bet is that the next step after chat is delegation. Instead of asking a chatbot for a summary, you set a goal and the agent keeps checking. That is a bigger change than it sounds, because the agent now has standing access to your accounts.
It also changes how to evaluate them. A chatbot is judged by answers. An agent is judged by whether its work is right when you are not looking, which is harder to measure and, so far, not independently measured at all.
How Does Claude Duet Work?
Duet is Anthropic's proactive assistant. According to The Neuron's coverage, Anthropic introduced it on October 21 as a research-preview waitlist for US Pro and Max subscribers, through dedicated mobile apps or the web. Team and Enterprise plans are not included.
It runs on Opus 5.5 and can use existing Claude memory and skills. A cloud-based virtual machine lets it keep going with browser tasks when your own devices are offline. It checks connected accounts every few hours, not continuously.
Anthropic says it can reach more than 6,000 marketplace connectors and plugins. Examples include drafting Gmail replies, checking unpaid invoices in QuickBooks or Stripe, preparing payroll in Gusto and following up with customers in HubSpot.
Approval is required before sending messages, making payments, signing agreements or deleting documents. Securities trading and transfers between financial accounts are prohibited even with permission.
How Does OpenAI Dots Work?
Dots is OpenAI's version. SiliconANGLE reported that OpenAI introduced it at its DevDay conference on September 29 as always-on agents inside ChatGPT, running on GPT-6 Astra. Each agent has its own cloud computer, which users can open at any time to see what it is doing.
Plugins reach more than 4,000 apps. Agents are available through ChatGPT on desktop, web and mobile, and through Slack and Teams, with context carried across channels. When no task is assigned, a dot runs "proactive research" with read-only connections, so it cannot send messages or change anything.
Safety checks include an "auto-review" step for actions that touch accounts or share information, and a monitoring system that can pause or stop an agent. OpenAI itself advises users to review consequential work, noting that agents can still make mistakes.
One agent comes with ChatGPT Pro or Business Premium at no extra cost, and agent usage does not count against plan allowances for the first month. Pro subscribers in the EEA, Switzerland and the UK are excluded. Enterprise customers can join a beta if a workspace administrator turns it on.
How Does Meta Muse for Small Business Work?
Muse for Small Business is the most vertical of the three. The Next Web reported that Meta announced it on September 29, 2026, as an expansion of the Muse personal agent that launched in the US and Canada on September 8.
Its connectors lean toward commerce and marketing: Shopify, Stripe, Canva and Intuit QuickBooks, plus analytics from Instagram business accounts and Facebook Pages, Meta ad accounts, Slack, Zoom, Notion, Figma and Dropbox. Owners can build custom connectors for services Meta does not yet support.
Muse cannot publish content, send messages or make purchases without the owner's sign-off. Most people can use it for free, with paid plans offering higher usage limits. Meta has not set separate pricing for the business version.
Two details deserve attention. Meta says about one in three early testers connected Muse to a business account. And Amazon removed Muse from its store, saying the agent had not identified itself and appeared to store customers' login credentials. Meta denies the credential claim.
Which Agent Costs the Least?
Muse is the cheapest to start, since most people can use it free. Dots includes one agent with an existing paid plan, and the first month does not count against allowances. Duet is available only to Pro and Max subscribers, so it requires an existing paid Claude plan.
The table below compares the launch terms as reported.
| Feature | Claude Duet | OpenAI Dots | Meta Muse for Small Business |
|---|---|---|---|
| Model | Opus 5.5 | GPT-6 Astra | Not specified in coverage |
| Access | US waitlist, Pro and Max | Pro and Business Premium | US and Canada |
| Cost | Counts against plan limits | One agent included | Mostly free, paid tiers for more use |
| Approval needed for | Sends, payments, signing, deleting | Auto-review of risky actions | Publishing, messages, purchases |
| Apps reached | 6,000+ connectors | 4,000+ via plugins | Commerce and marketing tools |
Treat the cost row with caution. None of the three has published typical usage figures for a full working day, and background checks can use up plan limits faster than a few chat messages would.
Which Agent Is Easiest to Supervise?
Dots offers the most visibility. Users can open its cloud computer at any time and watch what it is doing. Duet is weaker here: users cannot currently inspect its browser activity, which makes it hard to trace how it reached a recommendation, according to The Neuron.
That matters more than it sounds. If an agent drafts a payroll worksheet, you need to know what it looked at and what it left out. An omitted reimbursement can make a sheet look complete but be wrong.
Muse relies on approval gates. Because it cannot publish, send or buy without sign-off, the risk sits in what it presents for approval. A clean-looking request can hide a mistake as easily as a clean-looking report.
What Do the Launch Posts Leave Out?
All three leave out the number that matters most: how often the agent finishes a task correctly. Anthropic has not published task-completion rates for its workflows, and the Duet coverage notes no independent evaluations of accuracy were available. OpenAI says plainly that agents can still make mistakes.
They also leave out the cost of review. Frequent approval requests can add administrative work if the agent does not supply enough supporting information. The Neuron's analysis puts it as a simple test: the value depends on whether time saved gathering information exceeds time spent verifying recommendations.
Finally, they leave out failure modes. Duet checks accounts every few hours, so a last-minute change could go unnoticed until the next check. Anthropic also acknowledges that Duet can misinterpret conversations or task urgency.
How Should You Choose Between Them?
Choose based on where your work already lives. If your team is in ChatGPT, Slack and Teams, Dots fits with less friction. If you live in Claude and need browser tasks to continue offline, Duet is the natural fit once you clear the waitlist. If you run a shop on Shopify and keep books in QuickBooks, Muse is aimed squarely at you.
Whichever you pick, start small. Connect one account, give the agent one low-risk task and review every output for two weeks. Track how long review takes, because that number is your real cost.
The steps below make a sensible pilot.
- Pick one repeating task, such as unpaid invoice reminders or a morning inbox summary.
- Grant read-only access first, and add write access only after the outputs prove reliable.
- Keep approval on for anything involving money or customers.
- Log errors and time spent reviewing for two weeks.
- Expand only if the time saved clearly beats the review time.
Sandboxing and permissions matter here too. Our explainer on how AI coding agent sandboxes isolate risk covers principles that apply to any agent with standing access, and our comparison of decision models looks at the layer where an app lets a model decide. For what an agent actually remembers between runs, see our guide to AI agent memory systems.
The Bottom Line on Always-On Agents
The three agents are closer in design than in pricing and openness. Dots offers the most visibility, Muse the lowest cost of entry and Duet the tightest approval rules on money. None has independent proof that it does the work right.
At Model Drop, our editors will keep following these rollouts as pricing, usage limits and independent evaluations arrive. Your next step is to choose one low-risk recurring task, and test a single agent on it before granting access to anything involving payments.
