Anthropic 推出了对标 Meta Muse 和 OpenAI Dots 的主动式智能体应用 Claude Duet，首批面向美国 Pro 和 Max 订阅用户开启候补访问。该应用基于云端虚拟机运行 Claude Opus 5.5，每隔几小时在后台检查已连接的账户，涉及支付或敏感操作时需经双重确认并配合虚拟卡完成。
原文披露了主动式智能体应用 Claude Duet 的后台运行频率、底层模型及安全机制，读者可以据此评估其成本与实际可用性。
Anthropic rolls out Claude Duet, a proactive agent application that takes on Meta’s Muse and OpenAI’s Dots
Access starts with US Pro and Max subscribers on a waitlist.
what we know so far
Claude Duet checks connected accounts every few hours instead of continuously because constant checking would use up Pro and Max limits much faster, and those background checks draw on the same allowance as regular Claude chats.
Duet runs Claude Opus 5.5 on a cloud virtual machine with its own browser, but users can’t yet watch that browser, so they see results and approval prompts rather than each step.
> at Opus 5.5’s API list price a hypothetical check reading 40K tokens and writing 1.5K tokens would cost about $0.19.
6 such checks a day would come to about $34 a month at API rates, more than Pro’s $20 monthly price
Purchases need 2 separate approvals, with Duet confirming the item and Stripe’s Link wallet approving the spend, before Duet pays with a virtual card that keeps your real card number private.
Before Duet sends a message, makes a payment, signs a document or deletes anything, it shows exactly what it will do and waits for approval, yet trading securities and moving money between accounts stay blocked even with approval.
A new classifier and safety system prompt reject requests to share credit card numbers and government IDs, while health or personal-finance details go out only when the user explicitly tells Duet to send them.
Credential handling follows the Claude Managed Agents design, so Duet never directly sees passwords or sign-in tokens, and the auto mode classifier used in Claude Code screens its actions.
来源：Rohan Paul · x.com