Anthropic Launches Claude Duet, an Always-On Agent Powered by Opus 5.5

Anthropic today launched Claude Duet, a proactive AI agent that monitors connected apps, surfaces alerts, and executes tasks without waiting for the user to open a chat window. Available immediately via waitlist to US-based Claude Pro and Max subscribers on iOS, Android, and claude.ai, Duet is powered by Claude Opus 5.5 and runs on a dedicated cloud virtual machine — meaning it keeps working after you close your phone. The research preview carries no additional charge beyond existing subscription costs, though background checks and morning briefings consume plan usage quota.

The launch completes a six-week sprint by every major US AI laboratory into a newly formed product category. Meta's Muse agent arrived first on September 8 and 9, powered by Muse Spark. OpenAI unveiled Dots at DevDay on September 29, backed by GPT-6 Astra and more than 4,000 integrated apps. Google announced Gemini Spark at its May 2026 I/O developer conference, and Microsoft has separately outlined Autopilot capabilities for its enterprise Copilot suite. Anthropic, entering last among the four major labs, is staking its position on a specific claim: that Opus 5.5 is the first language model with enough judgment to proactively act on a user's behalf without constant instruction.

Whether that judgment claim holds under real-world conditions is the central question for a product that, by design, operates largely outside the user's view.

Anthropic Is the Fourth Lab to Enter the Proactive Agent Race in Six Weeks

Describing Claude Duet as a novel category would misread the competitive moment. By the time Duet opened its waitlist today, the "always-on AI agent" category had been contested for forty-three days. The question Anthropic is actually answering is not whether proactive agents are viable — Meta and OpenAI have already claimed they are — but whether its own model has the calibration to do the job better.

The connector ecosystem is where Anthropic makes its most concrete competitive claim. Claude Marketplace currently offers more than 6,000 connectors and plugins, surpassing OpenAI's disclosed count of 4,000+ apps for Dots. Both figures are company-reported and neither has been independently audited. The category of connectors also matters: Duet's launch highlights Gmail, Slack, Stripe, Gusto, DocuSign, QuickBooks, HubSpot, and Shopify as key integrations — suggesting its primary target is the small-business and independent professional segment rather than deep enterprise IT workflows.

The timing disadvantage cuts both ways. Anthropic enters after its competitors have established category expectations, meaning it faces comparison pressure from day one. But it also enters after those competitors have identified early friction points — from notification calibration to approval UI — that Duet can learn from before reaching general availability.

For existing Claude Pro and Max subscribers, Duet's most meaningful advantage may be architectural rather than feature-level: Duet inherits the user's existing Claude account state, including long-term memory, custom skills, and already-authorized third-party apps. A Pro subscriber who has spent months configuring Claude's memory and installing connectors gets a proactive agent that already knows them. A new user adopting Dots or Muse starts from a blank profile.

Read more: OpenAI launches parallel agents and GPT-6 Astra for ChatGPT Pro

Opus 5.5's Judgment Claim Underpins Duet's Proactive Architecture

Every proactive agent product launched in the last six weeks has framed the same engineering challenge identically: the model must decide when to interrupt the user and when to wait. Interrupt too often, and the agent becomes a nuisance. Interrupt too selectively, and users miss things they needed.

Mike Krieger, Instagram's co-founder and Anthropic's chief product officer, has pointed to Opus 5.5's calibration as the central reason Anthropic selected it to power Duet over lighter, faster options. According to Anthropic's launch materials, the company believes Opus 5.5 has the judgment needed to determine when to proactively act on a user's behalf — a decision threshold no previous model in its lineup was designed to handle autonomously. This is a company claim, not an independent evaluation.

Opus 5.5, released September 22, was positioned at the time as a significant step over its predecessor: 40% cheaper to run, 30% faster, and 85% less likely to attempt to circumvent usage limits — the last metric being particularly relevant for an agent that operates autonomously across authorized accounts. Anthropic described its performance as comparable to Fable 5.1, the company's safety-focused frontier model. The context window reaches one million tokens with a maximum output of 128,000 tokens per call, enabling Duet to ingest and reason across months of email, document, and transaction history in a single inference pass.

The pricing structure reflects Opus 5.5's positioning for extended autonomous use: $4 per million input tokens and $20 per million output tokens through the API. For an agent that regularly reads large bodies of connected data — a full inbox, a bookkeeping ledger, a CRM history — the input cost dominates. The company's 40% cost reduction claim relative to Opus 5 is directly relevant to the economics of running continuous background checks.

The judgment framing also signals what the model is being asked to do that earlier language models were not. A reactive assistant has a simple decision function: answer the question well. A proactive agent has a more complex one: determine, from unbounded context, which of hundreds of things it could surface is worth surfacing now. Getting that wrong in either direction — too noisy or too quiet — undermines the product entirely. The industry has not yet established benchmarks for agent calibration quality the way it has for reasoning or coding tasks, which means Anthropic's "sufficient judgment" framing cannot currently be independently verified against a standard test.

A Cloud VM Keeps Duet Working After You Close Your Phone

The infrastructure architecture separating Duet from a conventional AI assistant is a dedicated cloud virtual machine with its own browser. When Duet receives a task — or decides, via its own monitoring, that a task needs doing — it executes in Anthropic's cloud environment. The user's device does not need to be on.

This design choice has practical and technical implications that go beyond convenience. Practically, it enables Duet to run overnight processing, catch the 3 AM Slack message, compile the morning brief, and deliver it as a push notification before the user checks their phone — without draining battery or requiring a background process on a mobile operating system that is actively managed to limit such drain.

Technically, it means Duet's browser access is a fallback for services that lack a native connector. If a platform the user needs is not in the Claude Marketplace, Duet can attempt to interact with it through the browser directly, behaving as a human user would, using the visual interface rather than an API. This is both a capability and a risk surface: browser-based automation is inherently more brittle than native connector integrations and can fail when a website updates its layout or implements bot detection.

The safety architecture runs parallel to the execution environment. A dedicated safety classifier monitors Duet's activity and operates alongside a safety system prompt designed to prevent sensitive information from one connected account from leaking into context where another integration could access it. Anthropic states that Duet never accesses user passwords or login credentials directly; connections are established through OAuth-style authorization flows within Claude Marketplace.

The approval framework defines where Duet's autonomy ends. Four categories of action require explicit user confirmation before execution: sending messages, initiating payments (processed through Link, Stripe's digital wallet, or Shop Pay from Shopify), signing documents, and deleting content. A fifth category is permanently prohibited regardless of approval: trading securities or moving money between accounts. These permanent prohibitions map directly to the action types most likely to trigger financial regulation or consent law conflicts.

In Anthropic's launch demonstration, the agent's calibration manifested as a specific morning-brief format. A fictional coffee brand operator named Maya, with a Chicago store opening twenty-six days out, received a morning message synthesizing all active threads: forty-one items on the opening checklist reduced to two that needed her input that day — a health permit application with a $640 filing fee awaiting payment authorization, and a pitch deck slide requiring her judgment. The agent had already completed the permit form; it waited at the payment step. Duet had also identified that subscription cancellation rates had climbed from 2.6% to 3.4% following a shipping adjustment, and proposed a free-shipping-over-$40 threshold test. The owner approved it with two words. These are illustrative demo figures, not reported business outcomes, but the interaction architecture they depict — action surface, pause for approval, resume — reflects the actual product design.

Six Thousand Connectors and a Single Persistent Thread

The functional scope of what Duet can do is defined primarily by its connector layer. The Claude Marketplace integrations cover the core software stack of a modern small business or professional: Gmail for communication, Slack for team messaging, Stripe for payment processing, Gusto for payroll and HR, DocuSign for document signing, QuickBooks for accounting, HubSpot for customer relationship management, and Shopify for e-commerce. The 6,000+ connector figure presumably includes the long tail of specialized tools beyond these flagship integrations.

The single-thread model is a structural departure from conventional AI chat products. All interactions with Claude Duet — whether user-initiated or Duet-initiated — occur within one continuous conversation thread. This is architecturally significant: it means Duet maintains a unified context window across the history of its proactive activities and user responses, rather than treating each interaction as a fresh session. The practical consequence is that Duet builds what amounts to a running log of your working life across all connected apps — and the model can reference that accumulated context when making new judgments.

Three beta users described the product's early utility in Anthropic's launch materials. Jess Provencher, broker and owner of Pro Homes, said Duet identified a scheduling conflict — a property showing booked on the wrong date — before she had thought to check. South Florida restaurant operator Amin Ebrahimi, who manages roughly 110 employees across Shahs of Kabob locations, said Duet handling accounting, customer service, and marketing research eliminated his need to maintain parallel browser sessions for invoice review across multiple stores. Brand strategist Mya Dalton described Duet catching a conflict between a departing flight and a client call while preparing a pre-meeting brief, then offering to reschedule the call.

Two Anthropic employees described personal workflows from internal testing that were included in the company's launch materials. Boris Cherny, who leads Claude Code development, began dictating to-do lists to Duet verbally during walks, with Duet routing tasks without requiring him to stop or type. Chelsea Larsson, responsible for model and product behavior design, reported using Duet to offload her task list while handling childcare obligations, receiving a compiled results report at the end of the day.

These anecdotes describe Duet working as designed. They do not describe failures, and at this stage, failures are precisely what prospective users cannot yet evaluate.

What Duet Cannot Yet Do: Polling Gaps, Missing Microsoft 365, and Quota Consumption

Anthropic's own launch materials disclose several constraints that define the product's current boundaries more precisely than its capability list does.

Background checks occur at intervals measured in hours, not in real time. Anthropic's stated rationale is that continuous monitoring would exhaust a subscriber's plan usage limits; the poll cadence is a deliberate balance between responsiveness and resource consumption. The practical consequence is that time-sensitive developments arriving between polls — a bid acceptance, an escalating customer complaint, a time-stamped regulatory notification — may not surface until the next scheduled check. For the morning brief to reflect the prior night's actual events, those events must have occurred before the overnight poll ran.

Background checks and morning briefs count against the user's existing Pro or Max subscription quota. Duet does not operate on a separate allocation; every token it processes for monitoring, synthesis, and task preparation draws from the same budget as ordinary Claude conversations. Heavy users who regularly approach their plan limits may find Duet's background activity compressing the quota available for direct interactions.

Slack integration is described as functional but limited in scope. Microsoft 365 — encompassing Outlook, Teams, OneDrive, Word, and Excel — is not supported at launch. For professionals whose primary working environment is Microsoft's productivity suite, Duet's practical utility is substantially reduced at its current state. Given that Microsoft has its own Autopilot agent initiative within Copilot, the absence of Microsoft 365 integration may reflect both technical priority choices and competitive considerations.

Anthropic also acknowledges a qualitative limitation inherent to the proactive design: Duet may misread thread context or miscalibrate the urgency of specific items. A monitoring system that occasionally surfaces the wrong things — or misses something important — is worse than no monitoring at all for users who come to rely on it. The challenge is maintaining calibration across the full diversity of contexts that real users bring to it.

Users cannot observe Duet's cloud browser while it is performing tasks. The execution environment is opaque in real time; users see results, approval prompts, and morning brief summaries, but not the intermediate steps Duet takes within its VM. This is both a usability constraint and, for some users, a trust constraint: the system acts without a visible audit trail at execution time.

Read more: Anthropic Fable 5.5 spotted in Azure catalog as new model targets goal inference

Research Preview Scope and the Road to Team and Enterprise Adoption

Duet's launch today is explicitly scoped as a research preview. The waitlist format limits access velocity while Anthropic gathers calibration data from a controlled initial population of Pro and Max subscribers. No pricing beyond the existing subscription cost has been announced for the research preview period.

Team and Enterprise plans are not included in the current rollout. For Anthropic, this represents a significant portion of the commercial market: organizations with structured workflows, compliance requirements, and IT approval processes are precisely the segment where a proactive agent that spans accounting, HR, contracts, and payments would generate the most tangible business value. The absence of Team/Enterprise support at launch may reflect the additional safety, auditability, and administrative controls those deployments require, or simply a sequenced rollout prioritizing consumer feedback before organizational deployment.

Availability is currently limited to the United States. International expansion timing has not been announced.

Anthropic has indicated a forthcoming safety blog post will document the technical reasoning behind Duet's action boundaries and the safety classifier's design. That publication will matter: the permanent prohibitions and the approval-required categories represent judgment calls about where autonomous AI action should stop, and the company's public reasoning for those choices will be a meaningful reference document as regulators begin scrutinizing autonomous AI agents with payment and communication authority.

The proactive agent category now has four named products from four major laboratories, all launched or announced within five months. The connector ecosystem appears to be the early differentiator that is most legible to users and most defensible over time: a larger connector count creates network effects as more workflow integration points become available, and switching costs accumulate as users authorize more connections and accumulate agent memory specific to a given platform. Whether Duet's 6,000-connector lead translates into durable retention — or whether OpenAI's Dots closes the gap before general availability — depends on how quickly both companies expand their respective marketplaces and, ultimately, on whether Opus 5.5's judgment claim survives contact with the full range of what users actually hand it.