Claude Duet: Anthropic’s Agent Runs on Opus 5.5 [2026]

Anthropic opened a research preview of Claude Duet, a proactive agent running on Claude Opus 5.5 that messages users, works inside connected apps and keeps executing tasks on a cloud-based computer after the app is closed, rolling out via waitlist to Claude Pro and Max subscribers in the United States as of October 21, 2026. Anthropic said the agent asks for permission before sending a message, making a payment, signing a document or deleting anything, and acknowledged the preview "can still misjudge what's urgent." The launch makes Anthropic the third major US AI lab to ship a consumer-facing always-on agent, following Meta's Muse and OpenAI's Dots, introduced at DevDay three weeks earlier.

Anthropic has opened a research preview of Claude Duet , a new proactive AI agent that messages users, works inside connected apps, and keeps running tasks after they close the app. The move, reported by The Verge https://www.theverge.com/ai-artificial-intelligence/1015843/anthropic-claude-duet-ai-agent on October 21, 2026, puts Anthropic squarely into the fight for a new category of AI product: the “always-on” personal agent that Meta, OpenAI and Google have all raced to ship since late summer. Claude Duet is rolling out as a research preview to Claude Pro and Max subscribers in the United States, distributed through a waitlist rather than a flat-out launch. It runs on Claude Opus https://shattered.io/claude-opus-6-rumors-opus-5-5-cost-cut-2026/ 5.5 , the model Anthropic previously pitted against GPT-6 Astra on Terminal-Bench, and it is designed to behave less like a chatbot and more like a colleague who checks in, flags what needs attention, and asks before taking any action with real-world consequences. The announcement lands less than two months after Meta launched Muse and barely three weeks after OpenAI introduced its own persistent agent, Dots, at DevDay. Anthropic is now the third major US AI lab to ship a consumer-facing agent built to run unattended, and the timing, availability, and safety design all say something about how Anthropic thinks it can win a category it entered last. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F What Is Claude Duet, Exactly? Claude Duet is described by Anthropic as a research preview of Claude that users message the way they would message a person, rather than typing prompts into a chat window and waiting for a reply. The product ships as standalone apps for iOS and Android, alongside a web version at claude.ai, giving it the same multi-platform footprint that rival products have used to drive adoption. What separates Duet from the standard Claude chat experience is persistence. Instead of waiting for a prompt, the agent proactively surfaces what it judges needs the user’s attention, works through connected apps, and keeps operating on its own cloud-based computer even after the user has closed the app entirely. That last point is the technical crux of the “always-on agent” category: the work does not pause just because the device did. Anthropic’s own description, posted alongside the launch, frames the behavior directly: “It comes to you with what needs your attention, completes tasks you ask of it, keeps working when you step away, and checks with you before it acts on your behalf,” the company said in its official announcement https://x.com/claudeai/status/2112984669943968099 . That framing, task triage plus unattended execution plus a built-in checkpoint, is Anthropic’s answer to the question every lab shipping an agent product has had to confront this fall: how much autonomy is too much before users stop trusting it. The Approval Gate Is the Whole Pitch Anthropic has built explicit checkpoints into Duet’s workflow. The agent is designed to ask for permission before it sends a message on the user’s behalf, makes a payment, signs a document, or deletes anything. That is a narrower, more conservative set of guardrails than a general “ask before anything important” policy, and it signals that Anthropic is treating financial and irreversible actions as the bright line that triggers a human check-in, while leaving lower-stakes research, drafting, and monitoring tasks to run autonomously. Anthropic’s own account was candid about the preview’s limits. “Claude Duet is an early research preview, available through a waitlist on Pro and Max plans in the US. It can still misjudge what’s urgent, and we’ll keep tuning it based on what early users tell us,” the company said in the same launch statement. That kind of hedging is unusual for a product launch and reads as a deliberate attempt to set expectations low ahead of the inevitable stories about an agent that pinged someone about the wrong thing at the wrong time, or asked for approval on something trivial while missing something that mattered. The approval-first design also doubles as a liability shield. Every one of the four flagged actions, messaging, payments, signatures, deletions, is the category of mistake that turns into a support ticket, a news story, or a lawsuit if an AI agent gets it wrong unsupervised. Anthropic has spent the past year building a public safety narrative, including a proposed $1 billion AI safety bar that drew pushback from parts of the industry https://shattered.io/anthropic-ai-safety-bar-sparks-pushback-2026/ , and Duet’s design is consistent with a company trying to make caution part of its competitive identity rather than a compliance afterthought. Why Claude Opus 5.5 Is Doing the Heavy Lifting Anthropic built Duet around Claude Opus 5.5 specifically because of the judgment calls the product has to make constantly: what counts as urgent, what can wait, and when the agent needs to stop and ask. “We designed Claude Duet around Claude Opus 5.5, which gives it the judgment to decide what needs you, what can wait, and when to check with you first,” Anthropic said in a follow-up post explaining the model choice. That is a notable allocation of Anthropic’s flagship model to a free-feature research preview rather than reserving it purely for premium coding or enterprise workloads, where Opus 5.5 has already been positioned as a direct challenger to OpenAI’s GPT-6 Astra on agentic coding benchmarks https://shattered.io/claude-opus-5-5-vs-gpt-6-astra-terminal-bench-2026/ . It suggests Anthropic sees consumer-facing judgment and reasoning, not just raw coding throughput, as a core use case for its most capable model, and that the company is willing to spend inference budget on background agent work rather than ring-fencing Opus 5.5 for paying developers alone. Duet also inherits the user’s existing Claude account state. Anthropic says the agent “carries over the memory, skills, and connected apps from your Claude account, so there’s little to set up,” according to the same post. That continuity builds on Claude’s existing chat search and memory features https://support.claude.com/fr/articles/11817273-utilisez-la-recherche-de-chat-et-la-memoire-de-claude-pour-vous-appuyer-sur-le-contexte-precedent , which already let Claude recall prior conversations before Duet layers proactive action on top. That matters competitively: a user who has spent months teaching Claude their preferences, connecting tools, and building custom skills does not start from zero when they opt into the agent preview, which lowers the switching cost for existing subscribers and raises it for anyone tempted to jump to a rival platform that would require rebuilding that context from scratch. Who Can Actually Use Claude Duet Today Access is deliberately narrow. Claude Duet is available only to Claude Pro and Max subscribers in the United States, and even then only through a waitlist rather than an immediate rollout to every eligible account. Anthropic has not extended the preview to Team or Enterprise plans, though the company has said it plans to broaden availability later. For now, that leaves Anthropic’s business customers, the segment that generates the bulk of serious usage volume for most AI vendors, without access to the company’s most ambitious agent product. That staged rollout mirrors how Anthropic has handled previous capability launches, favoring a slow, US-first, subscriber-gated expansion over the kind of broad simultaneous release Meta used for Muse. It is a notably different go-to-market posture than the one OpenAI used for Dots, which launched already wired into more than 4,000 third-party applications. Anthropic appears to be optimizing for controlled feedback from a smaller, already-paying user base before it worries about matching the scale of its rivals’ integration ecosystems. No Price Tag, No New Bill Anthropic has not disclosed a standalone price for Claude Duet. Access comes bundled into existing Pro and Max subscriptions, meaning there is no new line item for the agent itself. Background agent activity does count against the usage limits tied to a subscriber’s existing plan, so heavy Duet use could, in theory, eat into the same quota a subscriber would otherwise spend on conversational Claude use. Anthropic has not published specifics on how much of a plan’s allowance a typical Duet session consumes. The Agent Wars: How Claude Duet Stacks Up Against Muse and Dots Claude Duet does not enter an empty field. Meta launched Muse in September, and the app reportedly climbed download charts fast, leaning on Meta’s existing distribution across Instagram, WhatsApp, and Facebook https://shattered.io/meta-muse-openai-dots-ai-agents-seo-2026/ . OpenAI followed within weeks, unveiling Dots at its DevDay event with a persistent cloud computer and browser of its own, plus task execution across a reported 4,000-plus connected applications. 9to5mac framed the launch https://9to5mac.com/2026/10/21/anthropic-takes-on-chatgpt-dots-and-meta-muse-with-claude-duet/ explicitly as Anthropic taking on both rivals at once. Google, meanwhile, has pushed its own Gemini-based personal-agent features, including a pricing structure that already shows wide gaps against Claude and ChatGPT tiers https://shattered.io/claude-vs-chatgpt-vs-gemini-price-gap-2026/ . Anthropic is therefore not creating the always-on agent category, it is the fourth major entrant into one that formed over roughly six weeks between early September and late October 2026. That compressed timeline is itself a story: four of the industry’s best-funded labs independently concluded, within the same two-month window, that a persistent, cross-app, proactive agent was the next product surface worth racing for. Coverage of the launch spread quickly beyond US outlets, too; Ukrainian tech outlet Mezha https://mezha.net/ua/news/mezha-bf84e06876b320771cf4a083 anthropic beretsya do/ picked up the news within hours, a sign of how closely international tech media now tracks every move in the agent race. Meta Muse: Distribution First, Depth Later Meta’s approach to Muse leaned on scale from day one. The app reportedly reached millions of downloads shortly after launch by riding Meta’s built-in audience, though analysts have cautioned that download counts alone say little about sustained engagement or completed tasks. Meta has continued iterating on pricing and features for the product, including a premium “Wonder” tier priced as high as $200 a month in some tests https://shattered.io/meta-muse-wonder-ai-pricing-2026/ , a sharp contrast with Claude Duet’s current no-extra-charge model. OpenAI’s Dots: Built for Scale Across Apps OpenAI’s Dots arrived with a wide integration net already cast, reportedly working across thousands of applications through DevDay-era partnerships. Some coverage has linked Dots to OpenAI’s GPT-6 Astra model, though that association has largely appeared in secondary reporting rather than a direct OpenAI technical disclosure. The direct head-to-head framing between the two products has already become its own storyline, with OpenAI explicitly positioning Dots as an answer to Meta’s early lead https://shattered.io/openai-dots-vs-meta-muse-ai-agent-2026/ in downloads. Google’s Gemini: The Quiet Incumbent Google has not framed a single product the way Anthropic, Meta, and OpenAI have branded Duet, Muse, and Dots, but Gemini’s agentic features have been expanding steadily across Google’s own surfaces, from search to productivity apps to connected devices. Google’s advantage is less a single flashy launch and more the sheer footprint of Gemini across products billions of people already use daily, a distribution base that none of the three newer agent products can currently match. Table: The Persistent-Agent Landscape, October 2026 | Product | Company | Status as of Oct. 21, 2026 | Underlying Model | Access | |---|---|---|---|---| | Claude Duet | Anthropic | Research preview, waitlist | Claude Opus 5.5 | Pro/Max subscribers, US only | | Muse | Meta | Public launch reported Sept. 2026 | Meta’s in-house models | Broad consumer rollout | | Dots | OpenAI | Public launch DevDay, Sept. 2026 | Reportedly linked to GPT-6 Astra | Connected to 4,000+ apps per reports | | Gemini agent features | | Ongoing rollout across products | Gemini family | Built into existing Google surfaces | The table underlines a simple point: Anthropic is launching the most restricted version of this product category, both in geography and in subscriber tier, while its three rivals opted for broader, faster rollouts. That restraint could read as caution or as a company still working out the kinks before it exposes Duet to the scale Meta and OpenAI have already absorbed. Anthropic’s Longer Agentic Build-Up Claude Duet is not Anthropic’s first step toward agentic products, it is the latest in a pattern. The company has steadily pushed Claude from a conversational assistant toward a tool that takes action, most visibly through Claude Code, its developer-focused coding agent that competes directly with tools like Codex CLI and OpenCode https://shattered.io/claude-code-codex-cli-opencode-update-4-days-2026/ , and through computer-use capabilities that let Claude operate software interfaces directly rather than just answering questions about them. Duet represents the consumer-facing extension of that same philosophy: an agent that does not just answer when asked but decides, within defined limits, what deserves attention and acts on it. Where Claude Code targets developers running structured engineering tasks, Duet targets a much broader audience managing messages, scheduling, payments, and day-to-day digital errands. The two products share a lineage, Opus-class judgment applied to real-world action, but they are aimed at entirely different users. What’s Still Missing: Price, Enterprise Access, and Global Reach Three gaps stand out in what Anthropic has disclosed so far. First, there is no confirmed standalone pricing for Claude Duet beyond its inclusion in existing Pro and Max subscriptions, so it remains unclear whether Anthropic will eventually charge a premium the way Meta has tested with Muse’s higher-end tier. Second, Team and Enterprise customers have no access yet, which excludes the segment most likely to generate high-volume, revenue-driving usage of an agent that can message contacts, handle approvals, and work across connected business tools. Third, availability is limited to the United States, leaving the rest of Anthropic’s global subscriber base, including markets where Claude already competes hard against ChatGPT and Gemini, without the feature entirely. Each of these gaps is a lever Anthropic can pull later to expand the product’s reach, and each is also a reminder that Duet is explicitly a research preview, not a finished release. How quickly Anthropic closes these gaps, particularly the Enterprise gap, will say a lot about whether the company sees Duet as a retention tool for existing subscribers or a genuine new revenue line. Table: Claude Duet at a Glance | Attribute | Detail | |---|---| | Announcement date | October 21, 2026 | | Product stage | Research preview | | Eligible plans | Claude Pro, Claude Max | | Geography | United States only | | Platforms | iOS app, Android app, claude.ai web | | Underlying model | Claude Opus 5.5 | | Access method | Waitlist | | Team/Enterprise access | Not yet available | | Actions requiring approval | Sending messages, payments, signing documents, deletions | | Additional cost | None disclosed; usage counts against existing plan limits | Market Impact: What a Fourth Agent Launch Signals The arrival of a fourth major persistent-agent product in roughly two months is itself a market signal. When Meta, OpenAI, Google, and now Anthropic all ship competing versions of the same basic idea inside a single quarter, it suggests internal roadmaps that were already converging well before any of the launches went public, likely driven by shared assumptions about where consumer AI usage is headed next: away from single-turn chat and toward software that manages ongoing tasks with minimal supervision. For Anthropic specifically, the stakes are different from its rivals. Meta and Google can subsidize agent experimentation with advertising and cloud revenue that dwarfs anything Claude generates. OpenAI has consumer scale through ChatGPT that Anthropic has never matched. Anthropic’s commercial strength has been concentrated in developer and enterprise API usage, which makes Duet’s consumer-first framing, and its exclusion of Enterprise customers at launch, a somewhat unusual sequencing choice. It suggests Anthropic wants direct, unfiltered feedback from engaged individual subscribers before it commits to supporting the kind of business workflows that would turn Duet into a serious enterprise product. There is also a trust dimension that goes beyond any single company. Agents that can message people, move money, and sign documents on a user’s behalf raise the stakes on every mistake considerably higher than a chatbot that gives a wrong answer in a conversation. Anthropic’s decision to gate exactly those four categories, messaging, payments, signatures, deletions, behind mandatory approval suggests the company is betting that visible caution, not raw capability, will be the differentiator that keeps subscribers comfortable handing over more of their digital lives to an agent. Historical Context: From Chatbots to Agents in Eighteen Months It is worth stepping back to see how fast this shift has happened. As recently as 2024, the dominant framing for consumer AI products was the chat interface: type a question, get an answer, start a new conversation for the next question. The agent category that Claude Duet, Muse, Dots, and Gemini’s features now occupy represents a fundamentally different interaction model, one where the AI initiates contact, operates across multiple apps without a human issuing each individual command, and persists across sessions rather than resetting with every new chat window. That shift tracks closely with the broader move toward agentic coding tools that Anthropic, OpenAI, and others have pushed in the developer space over the same period, Claude Code and its rivals being the clearest examples. Claude Duet effectively imports that same agentic pattern, plan, act, check in when needed, into a consumer product aimed at people who have never written a line of code. Whether that translation works as well for scheduling a dinner reservation or drafting a difficult email as it does for refactoring a function is the open question Anthropic’s research preview is explicitly designed to answer. Risks and Open Questions Several questions remain unresolved by Anthropic’s initial disclosures. The company has acknowledged the agent can misjudge urgency, but has not detailed how often that happens in practice, nor what recourse a user has when Duet acts on a mistaken judgment that falls outside the four explicitly gated categories. There is also no public detail yet on how Duet handles conflicting instructions across the multiple connected apps it touches, or how it prioritizes competing tasks when a user has delegated several things at once. Privacy is another open area. An agent that carries over memory, skills, and connected apps by design is also an agent that holds a denser, more persistent record of a user’s digital life than a conversational assistant typically accumulates. Anthropic has not published specifics on data retention for Duet’s background activity, how long task history is stored, or what controls subscribers have over deleting that accumulated context if they stop using the feature. Predictions: Where Claude Duet Goes From Here - Expect Anthropic to expand the waitlist and geography gradually rather than all at once, following the same staged pattern it has used for past capability rollouts, with international markets likely trailing the US by months rather than weeks. - Team and Enterprise access will likely arrive within the next two quarters, since enterprise usage is historically where Anthropic’s revenue strength lies, and a consumer-only agent leaves that advantage unused. - Pricing clarity should follow once usage data comes in from the research preview; a dedicated paid tier, similar to Meta’s premium Muse offering, is plausible if background-agent costs prove heavier than standard chat usage. - The four-way race between Claude Duet, Muse, Dots, and Gemini’s agent features will likely consolidate around integration depth and trust rather than raw feature lists, since all four products now offer a broadly similar core pitch of proactive, persistent, cross-app assistance. - Anthropic will probably publish safety and incident data on Duet’s approval-gate performance, given the company’s public positioning around AI safety, which would set a transparency benchmark its rivals have not yet matched for their own agent products. How This Compares to Claude’s Other Recent Launches Claude Duet’s research-preview framing is consistent with how Anthropic has handled other recent releases. The company has generally favored staged rollouts and conservative pricing moves over splashy, all-at-once launches, a pattern visible in how it introduced Claude Opus 5.5 itself earlier this year, positioning it against GPT-6 Astra on specific coding benchmarks rather than broad marketing claims. Duet extends that same cautious, benchmark-and-safety-first posture into a product category where Anthropic’s competitors have generally moved faster and louder. That contrast in pace, Meta and OpenAI shipping broad, fast, high-download launches versus Anthropic’s narrower, slower, subscriber-gated preview, may end up being the most consequential difference between the four products, more than any single feature comparison. Software categories defined by trust and sustained daily use, rather than one-time downloads, often reward the company that gets the safety and reliability story right over the one that gets there first. Frequently Asked Questions What is Claude Duet? Claude Duet is a proactive AI agent from Anthropic that messages users directly, works across connected apps, and continues operating on its own cloud computer even after the user closes the app, asking for approval before taking sensitive actions. Who can use Claude Duet right now? Access is limited to Claude Pro and Max subscribers in the United States, distributed through a waitlist. Team and Enterprise plans do not have access yet. What model powers Claude Duet? Claude Duet runs on Claude Opus 5.5, which Anthropic says was chosen specifically for its judgment in deciding what needs user attention and when to ask for approval. Does Claude Duet cost extra? Anthropic has not disclosed a separate price for Claude Duet. It is included with existing Pro and Max subscriptions, though background agent activity counts against a subscriber’s existing usage limits. What actions does Claude Duet need approval for? The agent is designed to ask before sending a message on the user’s behalf, making a payment, signing a document, or deleting anything. How is Claude Duet different from Meta’s Muse or OpenAI’s Dots? All three are persistent, proactive AI agents, but Claude Duet launched with a far narrower rollout, US-only and subscriber-gated through a waitlist, compared to the broader consumer launches Meta and OpenAI used for Muse and Dots. Can Claude Duet work on Team or Enterprise Claude plans? Not yet. Anthropic has said it plans to expand availability to those plans later, but the current research preview is limited to individual Pro and Max subscribers. Is Claude Duet available outside the United States? No. The research preview is currently limited to the United States, with no confirmed timeline for international expansion.