How it scores
What we love (and don't)
// PROS
- No extra charge. Included in existing Pro and Max plans during the preview
- Runs on Opus 5.5 on its own cloud computer
- 6,000+ connectors including Gmail, Slack, Google Calendar, QuickBooks, HubSpot
- Cloud computer keeps working after you close the app
- Asks before sending messages, making payments, signing documents, or deleting anything
- Memory, skills, and app connections carry over from your existing Claude account
// CONS
- Waitlist-only in the US. Team and Enterprise plans excluded
- You can't watch or take over its cloud browser
- Checks accounts at set intervals, so urgent items can wait until the next check
- Background checks count toward your existing plan limits
- Microsoft 365 not yet supported
- No phone calls or texts. It sends email from accounts you've connected, with your approval
- Anthropic has not said what it will charge after the preview ends
Anthropic launched Claude Duet on October 21, 2026, as a research preview. Access is still limited. Invitations go out in waves, and the pool of people who have actually used it for more than a few days is small. I am going to tell you what the product does and doesn’t do, what the few early impressions say, and whether the waitlist is worth joining. I am not going to dress up thin evidence as a full verdict.
The short version: the architecture is credible. The gaps are real. The price is right for now.
What Duet actually does
Anthropic built Duet to solve a specific problem: Claude is a chat tool. You open it, you type something, you get a response. Duet works the other way. It monitors your connected apps, sends you a morning summary, and reaches out when something needs your attention or when a task is finished.
It runs on Opus 5.5, on its own cloud computer. That second part matters. The cloud computer keeps working after you close the app. You can hand off a task in the morning and find the result waiting for you without having to leave a chat session open.
Tasks fall into a few types. Duet can flag inbox and calendar items, draft replies, check for unpaid invoices, prepare payroll for your review, and get contracts ready to send. Larger jobs, such as research briefs, financial models and slide decks, land in the Make tab as editable files. A Tasks tab tracks what is in progress and flags items that need your decision. Only you mark a task complete.
For sites with no connector, it uses its own cloud browser. You can task it by text or by recording a voice message.
Approval gates and safety
Duet asks before it does anything consequential. The support page lists the gate: it waits for your approval “before sending a message, making a payment, signing a document, or deleting anything.” Approvals and logins run in its own cloud computer. When a site asks for a password, you type it into a separate form that sends it to the browser, not the chat. Each purchase goes through a one-time card via Link or Shop Pay, and it asks before each one.
Some things are off-limits regardless: trading securities, moving money between accounts, accessing someone else’s account.
The support page acknowledges prompt injection as a known risk. Duet is built to treat what it reads in emails and web pages as information and to act on what you ask, but Anthropic is upfront that no defense is perfect.
This is the right posture for a preview. The transparency about what Duet cannot guarantee is more useful than a security marketing page would be.
Plans, pricing, and the missing piece
Duet is included in Pro and Max at no extra charge during the research preview.
| Plan | Price | Claude Duet |
|---|---|---|
| Free | $0 | Not included |
| Pro | $20/mo (or $17/mo annual) | Yes, waitlist |
| Max 5x | $100/mo | Yes, waitlist |
| Max 20x | $200/mo | Yes, waitlist |
| Team | $20/standard seat/mo annual | Not yet |
| Enterprise | $20/seat/mo annual + usage at API rates (20-seat minimum) | Not yet |
The missing piece is what happens after the preview. Anthropic has not said what Duet will cost once the research period ends. That is a real gap for anyone deciding whether to build a workflow around it now. For Pro subscribers at $20/month, the calculus is different than for someone considering a Max upgrade partly on Duet’s strength.
Duet draws from your existing plan limits at the same rate as chat. The morning brief and background checks count toward your five-hour and weekly allowances. Admitted users get one full limit reset, available until November 30, 2026.
Connectors
Duet works with more than 6,000 connectors and plugins. The support page names Gmail, Google Calendar, Google Drive, Slack, QuickBooks, HubSpot, Docusign, and Calendly, plus Stripe for unpaid-invoice checks.
Microsoft 365 is not yet supported. Slack works with limits. For everything else, the cloud browser fills the gap for sites without a dedicated connector.
The real limitations
The product page itself lists two: it may misread a thread or miscalibrate what’s urgent, and it can miss something that arrives between check-ins.
The support page adds more: no browser view or takeover, no view of the files on your computer, and no phone calls or texts.
The no-browser-view constraint is the one I keep coming back to. Duet uses its own cloud browser to complete web-based tasks, but you can’t watch that browser or take it over. The product page says you can see what Duet is doing and steer it, but steering by message is not the same as grabbing the wheel when a web form goes sideways. For anyone who has used computer-use agents, this is familiar.
Early impressions
Access is restricted enough that the pool of external users who have written about Duet is small. The one early roundup of first impressions I could find leans positive. Dan Shipper at Every, after using it for about two weeks, called it the first personal agent he would call a pro tool. Peter Yang called it the most proactive agent he has tried (Source). Phil Simon found Notion integration worked with almost no configuration.
The criticism follows predictable lines. Shipper found email drafts weaker than ChatGPT’s Dots. Claire Vo loved the design but said it never nudged her and felt slow. Flavio Copes called the no-browser-view constraint the product’s main weakness.
Anthropic’s own caveat on X is worth quoting directly: “It can still misjudge what’s urgent, and we’ll keep tuning it based on what early users tell us.”
That is an honest positioning for a research preview. The caveats are real, the feature set is real, and the product is early. The few reviews available are cautiously optimistic but thin on depth because most people don’t have access yet.
Who Duet is for
Anthropic’s stated target is “professionals and small business owners with more on their plates than they can track.” The product page examples (finding unpaid invoices, preparing payroll, getting contracts ready to send) read like small business back-office work.
The workflow Duet is best suited for: you have a recurring set of tasks that could be described clearly to another person, they involve apps Duet supports, and the acceptable failure mode is a review step, not an automatic commit. Reviewing a draft before sending is fine. Having Duet send directly without review is not a workflow to design around a research preview.
The workflow Duet is not suited for: you need Microsoft 365, you need real-time response to urgent items, or you’re on a Team or Enterprise plan.
Alternatives
The direct category comparison is ChatGPT Dots and Grok Bot. Dots is included in ChatGPT Pro (from $100/month) or Business Premium; it runs on GPT-6 Astra with its own cloud computer, and its proactive research uses read-only tools. Grok Bot comes with The X Subscription at Premium ($30/month), and its Bots run in parallel on one shared cloud computer.
The price comparison is stark. Grok Bot’s entry point is $30/month. Duet’s entry point is the Pro plan at $20/month, which is lower, but the two products are different in what they do. Grok Bot runs multiple parallel agents that can coordinate. Duet is a single persistent assistant that monitors and executes.
For background on Meta’s Muse, the other personal agent in this wave, see our Meta Muse launch post and the October 14 roundup that covers Muse Wonder’s new model tier.
For context on Claude itself, the underlying model and chat product, our Claude AI review covers the model’s strengths and weaknesses independently of Duet.
The honest take
The architecture is what you would expect from Anthropic: deliberate about safety and honest about limitations. The product page limitation list is not buried. I don’t expect to find a prompt injection warning in a help article, and the support page has one.
The gaps are real and Anthropic says so. Interval-based checks mean time-sensitive items can wait. No browser view means you can’t take over a web task mid-flight. Microsoft 365 is missing. Post-preview pricing is unknown.
If you’re on Pro or Max, joining the waitlist costs nothing and the preview includes a limit reset. The right approach is to get access, test it with one or two actual workflows, and decide based on what you find. A review of a research preview with limited external reports can tell you the shape of the product. Only using it tells you whether it fits your work.
If you’re on a lower plan, there is nothing to evaluate here. Team and Enterprise users are waiting for a different announcement.