Claude Duet launch

Claude Duet arrived about six weeks after Meta's Muse ^[1]and roughly three weeks after OpenAI's dots ^[2], turning a quiet corner of the AI market into a three-way sprint for the 'always-on personal agent' category almost overnight. The pricing spread is stark: Duet ships inside the existing $20/month Pro plan, Muse offers a free entry tier, and dots requires a Pro-tier subscription starting at $100/month ^[1][2]. That gap matters more than the feature list - OpenAI built dots on GPT-6 Astra and gated it behind a premium price point, while Anthropic chose to bundle Duet into its cheapest paid tier, betting that reach beats margin in a land-grab phase ^[2]. For now, Duet's reach is still narrow: the research preview is waitlist-only and restricted to US-based Pro and Max subscribers, with Team and Enterprise plans excluded entirely ^[3].

The headline safety claim is approval-gating - Duet is designed to check with the user before sending a message, making a payment, signing a document, or deleting content, and it never directly sees passwords ^[5]. But the more revealing detail is what sits a level below approval entirely: Anthropic outright prohibits Duet from trading securities or moving money between accounts, meaning the approval prompt never appears for those actions because they are disallowed by design, not merely gated behind a confirmation click ^[4]. That distinction - hard prohibition versus approval checkpoint - is the real engineering tell about which failure modes Anthropic considers too costly to leave to a confirmation dialog, even one with a human in the loop.

Anthropic's own Chief Product Officer Mike Krieger posted a first-party testimonial describing weeks of personal use - Duet drafted fireside-chat questions and watched Slack over a weekend - signaling genuine internal confidence in the product. Technical-audience reaction has been cooler: power users argue Duet is 'just another wrapper' around capabilities they already replicate with scheduled tasks and Claude Code, and sentiment on launch threads landed closer to skeptical curiosity than excitement. The gap isn't really a product-quality disagreement - it's an audience-fit one. Duet's value proposition is explicitly aimed at people who can't or won't stand up their own agent harness; for that group, a bundled $20/month proactive agent with 6,000-plus pre-built connectors is the whole pitch, not a limitation ^[5]. The clearest evidence that the underlying use case is real, rather than hype, comes from people already describing comparable proactive agents unblocking a stalled project while they were on vacation, triaging industry newsletters, or monitoring pull requests - the demand exists, the only open question is whether a packaged agent like Duet captures it better than a hand-rolled one.

Anthropic says Duet needed a frontier model specifically because the job is judgment, not throughput - deciding what's urgent, what can wait, which of 6,000 connectors to use, and when interrupting the user is worth the cost of their attention ^[8]. The architecture backs that framing: Duet runs inside a cloud virtual machine with its own browser so it can keep working while a user's device is offline, but it checks connected accounts every few hours rather than monitoring in real time ^[5][7]. That's a deliberate trade-off: Anthropic is choosing periodic judgment over constant polling, reasoning over a snapshot of what changed rather than reacting to every individual update as it happens.

Anthropic's own launch messaging concedes the preview 'can still misjudge what's urgent,' and independent coverage echoes that Duet can misread threads or assign the wrong priority to an issue ^[9]. Less visible but arguably more consequential: Anthropic has not published specifics on how long Duet retains background task history or what controls subscribers have to delete accumulated context from the accounts it monitors ^[4]. Layered on top is a usage-economics question - background monitoring and larger Duet tasks draw from the same Pro/Max allowance as every other Claude interaction, so a user who leans on Duet heavily could see less headroom for the rest of their Claude usage ^[6]. None of these are disqualifying, but together they're the real test of whether non-technical users grant Duet the account access it's asking for.