Claude Duet Launches at $20 a Month: Opus 5.5 Agents and a Usage Limit Trap

Anthropic's Claude Duet arrives at $20 a month — a third of what OpenAI charges for an agent

On October 21, 2026, Anthropic opened a research preview of Claude Duet, a standalone app where Claude acts as a proactive agent rather than a chat window you return to. It runs on Claude Opus 5.5 and ships as native apps for iOS and Android plus the web at claude.ai. Access is limited to Pro and Max subscribers in the United States, distributed through a waitlist in waves.

The feature list is competitive. The pricing is the news.

The category already had three entrants

Every major consumer lab now sells an always-on agent with its own cloud computer, and each picked a different door price:

Meta Muse, launched September 8 — free with a weekly allowance, then $20/month (Power, 500M tokens/week) or $100/month (Maximum, 3B tokens/week). U.S. and Canada only.
xAI Grok Bot, beta from August 11 — included with Cursor Pro at $20/month.
OpenAI Dots, launched September 29 at DevDay — one dot included with ChatGPT Pro at $100/month, $200 or $500, or a Business Premium seat. Not in Free, Go ($8) or Plus ($20). Excluded in the EU, UK and Switzerland.
Claude Duet, launched October 21 — included with Claude Pro at $20/month and Max, at no extra charge during the preview.

Four weeks ago the cheapest way to get an always-on agent from OpenAI was $100. Anthropic is now offering one on its $20 tier, and Meta will give you a limited version for nothing. OpenAI's $20 Plus subscribers still have no dot.

What Duet actually does

Anthropic designed it around three behaviors, and the company is unusually explicit about the trade-offs in each.

It briefs you, then goes quiet. Every morning Duet sends a short brief on the most important items across your connected apps: a hiring deadline, a contract redline, a permit fee awaiting payment, a delivery problem. It checks those apps every few hours through the day and messages you only when something needs a decision. Anthropic says it checks every few hours rather than continuously "because checking nonstop would use up your plan's limits much faster." You can tell it what you want to hear about and how often. It takes tasks of any size in one thread. A meeting booking and a financial model start in the same place, by text or by voice. It decides which apps to use, does the work in the background, and notifies you when it is finished or needs review. Larger outputs come back as files — a research brief, a forecast, a slide deck, a dashboard, a prototype, a web app — collected in a Make tab. A separate Tasks tab tracks ongoing work and flags what needs your attention; only you can mark a task complete. It stops before the decisions that matter. Duet reads, drafts and prepares on its own. Before it sends a message, makes a payment, signs a document or deletes anything, it shows exactly what it is about to do and waits. Two things are off-limits even with approval: trading securities, and moving money between accounts.

One design detail stands out. Duet runs on its own virtual machine with its own browser, so tasks continue after your laptop or phone is off. It carries over your Claude memory, skills and connectors on first launch, so setup is minimal. It can also browse sites that have no connector, using the cloud browser directly.

The connectors, and the checkout path

Duet can use the 6,000+ connectors and plugins in the Claude Marketplace. Anthropic's own examples:

Gmail and Slack — surface messages you missed, draft replies
QuickBooks and Stripe — find unpaid invoices, draft reminders
Gusto — prepare and run payroll with hours, bonuses and reimbursements filled in
LinkedIn — draft a job post, publish after you edit and approve
DocuSign — prepare agreements, send for review and signature
HubSpot — draft follow-ups from contact and deal history
Canva — turn a product launch into a social campaign
Shopify — set up, design and manage a store; a richer visual storefront editor is coming

Purchases go through Link, Stripe's digital wallet, and Shopify's Shop Pay. Link pays with a virtual card, so your real card number is never shared with the merchant. Signing in to sites with logins saved in 1Password is listed as coming soon.

That checkout path is worth noting next to OpenAI's Dots, which relies on a card saved at the merchant's site rather than a wallet partner. Meta Muse also uses Link and Shop Pay. On payments, Anthropic and Meta have converged; OpenAI has not.

The cost structure nobody should skim

Duet is included in Pro and Max during the preview at no additional charge. It is not free to run.

In Anthropic's own words, Duet "shares your plan's usage limits with the rest of Claude," and its background checks count toward them. The company warns users may hit those limits "sooner than with chat alone," and notes that larger jobs — a financial model, a deck — require more model turns and more background work than a normal conversation does. Newly admitted users get one full reset of their five-hour and weekly limits.

Anthropic has not published the token allowance behind the Pro and Max tiers, so a heavy user cannot budget Duet in advance. A polling agent that wakes every few hours burns tokens while you are asleep and idle. The practical approach is to watch consumption for a week before deciding whether the $20 tier is enough, because the failure mode is not a surprise bill — it is your regular Claude working time disappearing into background checks.

The same constraint applies to every product in this category. Meta publishes token allowances (100M a week free, 500M Power, 3B Maximum) but not a conversion from tokens to tasks. OpenAI says a dot's conversations don't count against ChatGPT usage, while work pushed into Codex or ChatGPT Work does, and it has not published the size of that allowance or the terms after the first month.

What isn't ready

Anthropic is direct about the gaps:

Not on Team or Enterprise plans yet — coming later.
U.S. only.
Weaker on personal errands than on work. Booking tickets is explicitly called out as a weak spot.
It can't show you its browser. You can see the output, not the browsing.
Separate app from the main Claude app for now; how the two merge will be decided by early-user feedback.
Microsoft 365 is not a connector yet.

Anthropic also acknowledges the judgment problem up front: early versions "might misread a thread or miscalibrate what's urgent." That is an honest statement about a product whose entire premise is deciding what deserves your attention.

How safety is wired

Anthropic says it will publish more detail on Duet's safety design in the coming days, but the mechanisms already named are specific:

Credential handling follows the Claude Managed Agents design, so Duet never directly sees your passwords or sign-in tokens.
Actions are approved by the auto mode classifier — the same safeguard that governs Claude and Claude Code — checking for steps you didn't ask for.
Prompt-injection and hidden-instruction defenses come from Anthropic's browser-use research.
Red teams and an external firm ran penetration tests, and Duet is covered by Anthropic's HackerOne bug bounty.
A new dedicated classifier and safety system prompt checks every message and guards against leaks of sensitive information. Requests to share credit card numbers or government identification are rejected outright. Health and personal financial information is allowed through only when you have explicitly instructed it to send.

That last item addresses the structural problem of an agent that reads email and sends messages: it must decide, constantly, where information is allowed to go. Splitting information into tiers and hard-blocking the top one is a more defensible answer than judging each case on merit.

What builders should take from it

Three things transfer regardless of region or plan.

First, the price of an always-on agent just collapsed. In September the answer from OpenAI was $100 a month. It is now $20 from Anthropic and $0 from Meta in limited form. Any product that charges for background autonomy needs a story for that. Second, shared usage limits are the honest architecture. Unlimited background polling does not exist at consumer prices, because the economics don't close. If you are designing a subscription with agentic behavior, meter the background work separately or expect your users to discover the ceiling in week one. Third, the two-tier permission model is a template. Approve-before-send, approve-before-pay, approve-before-sign, approve-before-delete — and never-trade-securities, never-move-money-between-accounts even with approval. When you hand an agent authority in your own system, that structure is close to the minimum viable policy.

Duet is not available in Japan or Europe, and the $20 door is a research preview that will change. But the pattern it sets — cheap entry, metered background work, explicit refusal categories — is what the rest of this market will be built on for the next year.