On October 21, 2026, Anthropic released Claude Duet, a proactive agent built on Opus 5.5 and positioned for professionals and small business owners. It sends a morning briefing drawn from connected email, calendar, banking and enterprise systems, and raises decisions that need the user's attention. It runs in its own cloud virtual machine with a browser, so it keeps working after the user closes the laptop. It must obtain approval before sending messages, making payments, signing documents or deleting data, and it will not execute securities trades or transfers between accounts even with authorization.
The research preview is limited to U.S. Pro and Max subscribers, included in the $20 per month tier and sharing usage limits with the rest of Claude. OpenAI's Dots sits in the $100 per month tier. Meta's Muse has a free tier and passed 14 million installs by October 21.
Judgment is the product and the attack surface
Anthropic makes judgment the core differentiator. In the launch demo, Duet tells a coffee shop owner that a Chicago location opens in 26 days and the health permit has not been filed — the application is already drafted. The $410,000 lease renewal, it says, can wait until Monday. Mike Krieger, co-lead of Anthropic Labs, describes the design goal as knowing "when to step in and when to act on your behalf."
Anthropic's own documentation concedes that Duet may "misread conversations or fail to judge what is urgent." That is not boilerplate caution. Duet checks in every few hours rather than continuously, and Anthropic's stated reason is practical: constant checking would burn through the user's quota. Low-frequency checking means two things. It can miss what happens between checks. And every time it does surface, it carries more weight.
A noisy agent trains users to ignore it. An agent that appears once or twice a day does not get ignored after a wrong call — it gets doubted. The low-frequency, high-judgment design trades restraint for trust. The fragility is that trust takes many correct calls to build and one false "urgent" to break.
Lenny's Newsletter called Duet the "most polished agent launch" while judging it "not proactive enough." That contradiction is the point. Anthropic lowered proactivity to avoid being annoying, but the category's promise is that the agent actually acts first. Taste is the product, and taste has no standard answer.
The lead came from Meta, and the form converged in 71 days
Karan Ahuja was in product management at Meta Superintelligence Labs until July, joined Anthropic in August, and led Duet. On X, he said the team built it for "professionals and small business owners juggling many things at work and at home." Meta's Alexandr Wang posted a meme implying Duet resembles Muse.
The personnel detail is not gossip. It shows how fast the proactive agent form is converging. From Grok Bot entering beta on August 11 to Meta Muse on September 8, OpenAI Dots on September 29, and Claude Duet on October 21, four companies shipped nearly identical products in 71 days: cloud VM, background checks, proactive notifications, approval before execution. Daily.dev's verdict: "the winner is whoever's least annoying."
When one person can move from Meta to Anthropic in two months and lead the same category, the design space has narrowed sharply. Competition is no longer about building a different product. It is about building the same product with fewer errors and fewer interruptions. Ahuja brought not code but a judgment about what a proactive agent should look like — and Meta's Muse and Anthropic's Duet share that same origin.
$20 vs $100 is two product philosophies
Duet sits in the $20 Pro tier. Dots sits in the $100-and-up tier. The gap reveals two bets.
Anthropic is betting proactive agents should be a mass product. Twenty dollars is the entry tier; users do not pay a premium for proactivity. OpenAI is betting the compute cost of proactivity needs a high price to cover. Dots runs background probing, continuous loops and 4,000-plus app integrations, and $100 is the cost-backward result.
Duet's $20 comes with a clause: it shares usage limits with the rest of Claude, and background tasks count. Digital Today notes that this means quota "may drain faster than with chat alone." What users will feel is not "$20 gets me a proactive agent" but "how much of my $20 did Duet eat." Anthropic chose to acquire first and handle friction later. That works in a growth phase. When users discover their coding quota is being consumed by an agent that pushes three items every morning, retention will answer.
P.S. In the launch video, Krieger told Duet to page him over the weekend only if something was on fire. It paged him once — proof the product works, and the exact reason users will never know whether that one page was real.
Frequently Asked Questions
Q: What is Claude Duet?
A: A proactive agent from Anthropic built on Opus 5.5, released October 21, 2026. It sends morning briefings drawn from connected email, calendar, banking and enterprise systems, runs in its own cloud VM, and requires approval before sending messages, making payments or signing documents.
Q: How is it different from OpenAI Dots and Meta Muse?
A: Duet checks in every few hours rather than continuously, and Anthropic frames the differentiator as judgment — deciding what needs the user and what can wait. Dots runs continuous background probing at $100 per month; Muse has a free tier and passed 14 million installs.
Q: Why does the low-frequency design matter?
A: Because every appearance carries more weight. A noisy agent trains users to ignore it. An agent that surfaces once or twice a day gets doubted after a wrong call. Anthropic's stated reason for low frequency is quota cost, but the product consequence is that trust is built and broken at a lower frequency.
Q: Who leads Duet?
A: Karan Ahuja, who was in product management at Meta Superintelligence Labs until July 2026 and joined Anthropic in August. Meta's Alexandr Wang posted a meme implying Duet resembles Muse, highlighting how fast the category has converged.
Q: How much does Duet cost?
A: It is included in the $20 per month Pro tier and shares usage limits with the rest of Claude. Background tasks count toward quota, so heavy Duet use can drain limits faster than chat alone.
Q: Why is the $20 vs $100 gap significant?
A: It reflects two bets. Anthropic treats proactive agents as a mass product at entry pricing. OpenAI treats proactivity as compute-expensive and prices it at $100 and up. Which model works will be answered by retention, not launch numbers.
