TL;DR
- Anthropic launched Claude Duet on October 21, 2026: a standalone proactive agent on Opus 5.5, limited to US Pro and Max subscribers via waitlist.
- It costs nothing extra, but it draws on the same usage limits as the rest of Claude — so Anthropic throttles it to checks “every few hours” instead of continuously, and says so openly.
- OpenAI’s Dots went the other way: a separate allowance, sold at $100 and $500 a month and $125 per seat on Business Premium.
- That decision — who eats the inference cost — now shapes the product more than any benchmark does.
Introduction
Three weeks separate OpenAI’s Dots from Claude Duet. Six weeks separate Meta’s Muse from Duet. In about two months, every major consumer lab has shipped an agent with its own cloud computer that works while you sleep, and the feature lists are now nearly identical.
What differs, and what almost nobody is pricing correctly, is the meter.
(Source : Reuters — Anthropic unveils Claude Duet AI agent app)
What Duet actually does
Duet lives in its own app — iOS, Android, claude.ai — separate from the main Claude app, and inherits your memory, skills and connectors, so setup is a matter of permissions. On Opus 5.5 it reaches the Marketplace’s 6,000-plus connectors, from QuickBooks and Stripe for unpaid invoices to Gusto for payroll prep, Docusign and Shopify, plus Link and Shop Pay for approved purchases with a virtual card.
The interaction model is the real departure. It sends a morning brief, then goes quiet, re-checking connected apps and messaging you only when something needs a decision: a $640 permit fee awaiting one tap, a contract redline, a dashboard rebuilt at 5:30 each morning.
(Source : Anthropic — Claude Duet)
The choice that shapes the product
Duet is included in Pro ($20 a month, $17 annual) and Max (from $100, at 5x or 20x Pro usage) with no separate subscription during the preview. That reads as generous. It is really a cost-allocation decision, and the bill arrives as latency in your own work.
Background checks count toward your plan limits, and a deck or a financial model consumes far more model turns than a chat reply. Anthropic’s stated reason for checking every few hours rather than nonstop: constant checking “would use up your plan’s limits much faster.” The throttle is not a design preference. It is the arithmetic of bundling.
OpenAI chose differently. Conversations with your dot do not count against ChatGPT usage limits, and the tier including one is priced accordingly: Pro at $100 a month, Pro 500 at $500, Business Premium at $125 per seat. xAI’s Grok Bot ships with its own allowance, separate from Grok and Cursor plans. Three vendors, three meters, and only one of them makes “always-on” a promise the product can actually keep.
(Source : SaaS City — Claude Duet: proactive agent research preview)
Anthropic also capped the blast radius: no Team or Enterprise plans, no API. The budgets that matter most cannot procure it yet.
The control envelope has converged
More telling than feature parity is safety parity. Two labs working independently, five weeks apart, drew the same box:
- A separate cloud computer that is not your laptop.
- Approval before send, pay, sign or delete — Duet shows the action and waits.
- Classifiers on the action itself: Anthropic reuses the auto mode classifier from Claude and Claude Code, plus a new sensitive-data classifier that refuses to transmit card numbers and government IDs.
- Hard prohibitions that survive your approval: no trading securities, no moving money between accounts.
That last item is the one to study. A vendor shipping an agent able to prepare and run payroll announces in the same breath that the agent will not move funds — because the liability of an autonomous transfer is not worth the demo. Agent capability is now paced by insurance, not by models. Duet never sees your passwords or sign-in tokens, and it was red-teamed, externally penetration-tested and covered by Anthropic’s bug bounty.
(Source : ExplainX — Claude Duet: proactive agent research preview)
What is still missing
Two gaps matter more than the geography. First, Duet cannot show you its browser yet — you get results without a step-by-step trace, which is a blocking gap for anyone who needs an audit trail before trusting an agent with invoices. Second, it is a second app beside Claude and Cowork rather than inside them, fragmenting the memory story the launch leans on. It builds on Anthropic’s agent stack rollout and on the Opus 5.5 price war that set the token economics underneath.
For a small US business on Stripe, QuickBooks and Gusto, Duet is the first proactive agent whose context arrives pre-loaded — and the first whose cost you feel mid-workday. Our OpenAI Dots breakdown covers the counter-model.
FAQ
Is Claude Duet free? No separate subscription during the preview. It is included in Pro and Max but consumes the same usage limits as chat and Claude Code.
Do background checks count against my limits? Yes. They draw down the plan allowance, which is exactly why Anthropic schedules them every few hours instead of continuously.
Can Duet pay for things or trade stocks? It can buy through Link and Shop Pay with approval. It cannot trade securities or move money between accounts — hard limits that survive approval.
When is it available outside the US, or on Team plans? Anthropic has not said. Launch is US-only, Pro and Max, waitlist, no API, no Team or Enterprise access.
How is Duet different from Cowork and Claude Code? Cowork works on files and apps on your desktop; Claude Code works in your repository and terminal. Duet works in its own cloud computer and initiates contact when something needs you.