Anthropic opened a research preview of Claude Duet on Tuesday, a proactive AI agent the company built to sit on top of its Claude assistant and work through tasks with little day-to-day prompting. The move, first detailed by The Verge, puts Anthropic in direct competition with Meta’s Muse family of agents, which Meta has been expanding aggressively since early September. Claude Duet is live today, October 21, 2026, for Claude Pro and Max subscribers in the United States who join a waitlist, running on Anthropic’s Opus 5.5 model.
The launch lands in the middle of what has quickly become the defining fight in consumer AI: not which chatbot answers questions best, but which agent can be trusted to act on a person’s behalf while they are doing something else. Anthropic, Meta, and OpenAI are now all racing toward the same idea from different angles, and Claude Duet is Anthropic’s clearest answer yet to a question the industry has been circling for more than a year.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Claude Duet Actually Does
Claude Duet is built around a single premise: most of what a person needs from an assistant is not a conversation, it is follow-through. According to Anthropic’s own product page, cited by Unite.AI, Claude Duet identifies what needs a user’s attention, takes on tasks of varying size, keeps working while the person is away from their device, and checks in before acting when a decision carries real consequences. That last part is the design choice Anthropic is leaning on hardest: autonomy that pauses for a human sign-off rather than autonomy that simply executes.
In practice, Anthropic has described the agent handling morning briefings that scan a user’s connected apps and surface what actually matters, monitoring inboxes and calendars for things that need a response, setting up meetings, building dashboards, drafting replies to messages, and putting together briefs or slide decks. On the administrative side, Claude Duet is described as capable of finding unpaid invoices and drafting reminders, preparing payroll data for a manager’s review, and getting contracts ready to send out. None of this requires the user to open a new app or start a fresh conversation from scratch, since Claude Duet carries over memory, skills, and connected apps from the main Claude app into the same thread.
Opus 5.5 Is Doing the Heavy Lifting
Claude Duet runs on Opus 5.5, the model Anthropic has positioned as its top-tier reasoning system this year. Opus 5.5 recently topped an independent LLM leaderboard comparison against GPT-6.1 Sol and Gemini 4 Argon, a result our earlier coverage of the LLM leaderboard standings broke down in detail, and it also set the pace on Terminal-Bench 4.0, a benchmark we covered when Opus 5.5 hit 65.15% against rival models on long-horizon terminal tasks. Choosing Opus 5.5 rather than a lighter, cheaper model for Claude Duet signals that Anthropic is treating proactive agent work as a reasoning-heavy problem, not a simple scheduling layer bolted onto a chatbot.
That decision has a cost implication that Anthropic has not detailed publicly. Running a top-tier reasoning model continuously in the background, across morning briefs, inbox monitoring, and multi-step task execution, is far more compute-intensive than a single chat response. Anthropic has not published a separate price for Claude Duet beyond the existing Pro and Max subscription tiers, and no official figure has surfaced confirming whether that will change once the research preview ends.
Who Can Actually Use It Right Now
Access is narrow by design. Claude Duet is restricted to Claude Pro and Max subscribers located in the United States, and even within that group, access runs through a waitlist rather than an immediate rollout. Anthropic has not extended the preview to Team or Enterprise customers, a notable gap given how much of the described functionality, invoice tracking, payroll prep, contract drafting, reads like it was built with small business operations in mind rather than individual consumers managing their personal inbox.
Users can reach Claude Duet through standalone iOS and Android apps or through the existing web interface at claude.ai. That multi-platform approach mirrors how Anthropic has rolled out Claude more broadly, rather than treating the agent as a separate, siloed product line.
Why a Waitlist Instead of a Full Launch
Anthropic’s choice to call Claude Duet a “research preview” rather than a general release fits a pattern the company has followed with other agentic features. Giving an AI system standing permission to act across a person’s calendar, inbox, and financial documents raises failure modes that a simple question-answering chatbot never has to face: a misread email, a miscategorized invoice, or a scheduling conflict handled incorrectly can cause real-world problems rather than just an awkward response. A staged rollout to paying subscribers who have already opted into Anthropic’s ecosystem gives the company a controlled population to watch for exactly those failures before expanding access.
Meta’s Muse: The Agent Anthropic Is Racing Against
Claude Duet is not entering an empty field. Meta has spent the past seven weeks building out Muse into what it now calls a frontier-class agent platform, and the timing of Anthropic’s launch, one day after Meta described Muse as having reached “the frontier,” is hard to read as coincidence.
From Muse Spark to Muse Wonder
Meta’s agent push started gaining momentum on September 2, 2026, when the company shipped Muse Spark 1.3, which it described as delivering stronger performance across agentic and coding tasks. Six days later, on September 8, Meta introduced Muse itself, calling it the “world’s first personal AI agent built for everyone.” At Meta Connect 2026 later that month, the company said it would extend Muse beyond phones and desktops to its smart glasses line, giving the agent a presence throughout a user’s day rather than confining it to a screen.
The bigger jump came on October 12, 2026, when Meta introduced Muse Wonder and Muse Wonder Full, which the company described, per Meta’s own newsroom post, as its most capable models to date, built specifically for personal and professional agentic tasks ranging from navigating an insurance claim to running multi-step work projects. Muse Wonder is available through the Meta Model API and Muse Code, while Muse Wonder Full is reserved for Muse Code subscribers. CNBC reported that Meta is positioning Muse Wonder as the most advanced option currently available for powering digital assistants, a claim that puts it in direct rhetorical competition with what Anthropic is now shipping in Claude Duet.
Meta has described Muse’s underlying approach in terms that go further than what Anthropic has said publicly about Claude Duet. According to a Meta statement reported by the Associated Press, “Once a person shares a goal with Muse, it helps them develop a personalized plan and coordinate their time and resources, then advances the work on its own. It can open a browser, fill out forms, and negotiate on their behalf.” That description covers more autonomous action, browsing, form-filling, negotiation, than anything Anthropic has attributed to Claude Duet so far, where the emphasis stays on monitoring, drafting, and preparing work for a human to approve.
Claude Duet vs Meta Muse Wonder: A Feature Comparison
Laid side by side, the two products reveal two different bets about what users want from a proactive agent in late 2026.
| Feature | Claude Duet (Anthropic) | Muse Wonder / Muse Wonder Full (Meta) |
|---|---|---|
| Status as of Oct. 21, 2026 | Research preview | Available via Meta Model API and Muse Code |
| Underlying model | Opus 5.5 | Muse Wonder / Muse Wonder Full |
| Who can access it | Claude Pro and Max subscribers, US only, via waitlist | Muse app, Meta AI surfaces, Muse Code subscribers |
| Platforms | iOS, Android, web (claude.ai) | Mobile, desktop, and Meta’s smart glasses |
| Core behavior | Monitors connected apps, drafts and prepares tasks, pauses for approval before acting | Builds a plan toward a stated goal and advances it independently, including browser actions and form-filling |
| Team/Enterprise access | Not yet included | Not applicable; consumer-led with a developer-facing Muse Code tier |
| Memory handling | Carries over memory, skills, and connected apps from the Claude app | Persistent memory, goals, and tracked artifacts across sessions |
The pattern in that table is consistent: Meta is building Muse to act further out on its own, while Anthropic is building Claude Duet to prepare work and wait for a nod. Both approaches carry risk, just different kinds. An agent that negotiates and fills out forms on a user’s behalf can save more time, but a wrong assumption travels further before anyone notices it. An agent that pauses for approval is safer by construction, but only if users actually read what it is asking them to approve rather than clicking through on autopilot.
The Wider Proactive Agent Race
Anthropic and Meta are not alone in this race, and the broader shift has already reshaped how people search for and interact with AI tools. Our earlier coverage of how OpenAI’s Dots and Meta Muse are reshaping search behavior showed that the arrival of agent-style assistants is already changing what people type into Google, since fewer queries look like questions and more look like task instructions. We also compared pricing and plan structures across the field in our breakdown of ChatGPT’s Dots, Meta Muse, and Claude’s agent tier, where the gap between free and paid access to proactive features was already one of the sharpest divides in the consumer AI market, before Claude Duet even existed.
What makes this moment distinct is that all three major labs are now shipping some version of “an agent that works while you’re not looking,” rather than treating that as a future roadmap item. That is a meaningful shift from even six months ago, when most agent demos were confined to coding tasks. Anthropic’s own developer tooling reflects that shift too; our setup walkthrough for the Claude Agent SDK running on Sonnet 5.5 showed how far the building blocks for autonomous task execution have already spread into Anthropic’s developer-facing products, ahead of a consumer-facing agent like Claude Duet catching up.
A Timeline of the 2026 Proactive Agent Push
| Date | Milestone | Company |
|---|---|---|
| Sept. 2, 2026 | Muse Spark 1.3 ships with improved agentic and coding performance | Meta |
| Sept. 8, 2026 | Muse launches, described as the “world’s first personal AI agent built for everyone” | Meta |
| Late Sept. 2026 | Meta Connect 2026 announces Muse expanding to smart glasses | Meta |
| Oct. 12, 2026 | Muse Wonder and Muse Wonder Full debut as Meta’s most capable agentic models to date | Meta |
| Oct. 21, 2026 | Claude Duet enters research preview, powered by Opus 5.5 | Anthropic |
Seven weeks separate Meta’s first Muse release and Anthropic’s answer, which is a tight turnaround for a company that has historically moved more cautiously than its two biggest rivals on consumer-facing agent features. That pace tells its own story: Anthropic is not willing to let proactive agents become a category Meta and OpenAI define without a direct competitor on the field.
Market Impact: What Investors and Enterprises Should Watch
For Anthropic, Claude Duet is as much a retention play as a growth play. Pro and Max subscribers already pay for Claude; giving them a reason to keep paying, by making the assistant proactive rather than purely reactive, raises the cost of switching to a competitor. The fact that Team and Enterprise plans are excluded from this preview suggests Anthropic wants to test the feature against individual behavior patterns before exposing it to the more complex permission structures and compliance requirements that come with business accounts managing shared calendars, shared inboxes, and financial systems across multiple employees.
Meta’s calculus is different. Muse is folded into a free, ad-supported ecosystem with billions of existing users, while Muse Wonder and Muse Wonder Full are also offered as developer-facing models through the Meta Model API and the separate Muse Code subscription. That gives Meta two revenue paths at once: consumer engagement data from free Muse usage, and a direct subscription product for developers who want to build on Muse Wonder. Fortune reported that Meta CEO Mark Zuckerberg told the Financial Times the company has reached “the frontier” with Muse Wonder and Muse Wonder Full, citing Meta’s own published benchmarks showing the models outperforming rivals including OpenAI’s GPT-6 and Anthropic’s models. Self-published benchmarks from any lab should be read with some skepticism until independent evaluations catch up, but the claim itself signals how central agent performance has become to Meta’s AI messaging.
Enterprise Safety Scrutiny Is Already Building
Anthropic’s cautious rollout also arrives against a backdrop of rising scrutiny over AI safety commitments industry-wide. Our earlier reporting on Anthropic’s own proposed $1 billion AI safety threshold showed that the company has been pushing for stricter industry-wide safety bars even as it ships more autonomous features of its own, a tension that is likely to follow Claude Duet as it moves from a US-only waitlist toward a wider release.
Historical Context: How We Got From Chatbots to Agents That Work Alone
It is worth remembering how recent all of this is. As recently as early 2025, the dominant AI product category was still the single-turn chatbot: a user asks a question, the model answers, the conversation ends. Agentic coding tools were the first place autonomy crept in at scale, letting a model write, run, and fix code across multiple steps without a human approving every line. Meta’s Chief AI Officer Alexandr Wang captured that shift bluntly, telling reporters, per NBC News, “It was only about six months ago that having a personal agent running on your own computer was the sign of someone who is completely AI-pilled and thinking nonstop about agents. But for that small group of nerds … it looked like the beginning of personal superintelligence. And today with Muse, we’ve brought that magic to everyone.”
What changed between those early coding agents and products like Muse and Claude Duet is scope. Coding agents operate inside a bounded, text-based environment with clear success criteria. A personal productivity agent has to operate across email, calendars, financial documents, and messaging apps, each with its own failure modes and each carrying real consequences for a wrong move. That expansion in scope is exactly why Anthropic’s approval-gated design and Meta’s more autonomous, goal-driven design represent two genuinely different philosophies rather than cosmetic branding differences.
What People Who Study AI Agents Are Saying
Meta’s own leadership has been the most publicly vocal about where this category is headed. Mark Zuckerberg described his long-term vision for Muse this way, as reported by the Associated Press: “Everyone will have an exceptionally capable personal agent that understands you, your goals, and everything you care about. Your agent will work 24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more.” That framing goes well beyond anything Anthropic has said about Claude Duet, which the company has described in narrower, task-completion terms rather than as a life-management system.
Alexandr Wang has also spoken to the practical side of unattended agent work, telling CNN, “You can have it help you run your small business, or just get work done for you. You can walk away from your computer, and it keeps working for you on all the jobs you line up.” That description tracks closely with what Anthropic says Claude Duet can do for invoices, payroll prep, and contracts, even though the two companies frame the underlying trust model differently.
On the safety side, Wang struck a different tone when discussing what will separate winners from losers in the agent race, telling NBC News, “My view is that trust and alignment are quickly becoming the most important capabilities that will differentiate agents and models. Any lab that doesn’t focus on alignment will fall behind.” That statement, coming from Meta’s own AI chief, lines up almost exactly with the approval-gated caution Anthropic has built into Claude Duet, even if the two companies arrived at that conclusion through very different product designs.
Five Predictions for the Next Six Months
- Expect Anthropic to expand Claude Duet’s waitlist to additional countries once the US research preview surfaces enough usage data to refine the approval-gating logic.
- Team and Enterprise access is likely to follow within a few quarters rather than years, since the invoice, payroll, and contract use cases Anthropic has already described are clearly aimed at small-business operators.
- Meta will keep pushing Muse into hardware, building on its Connect 2026 commitment to bring the agent to smart glasses, which gives Meta a distribution channel Anthropic and OpenAI currently lack.
- OpenAI is unlikely to sit still; given how our coverage of Dots versus Muse already tracked rapid feature parity moves between the two companies, a comparable proactive-agent push from OpenAI inside the next two quarters would not be surprising.
- Independent benchmarking of Muse Wonder’s claimed frontier-level performance will become a flashpoint, since Meta’s self-published comparisons against GPT-6 and Anthropic’s models have not yet been verified by a neutral third party.
The Trust Question Nobody Has Fully Answered
Both Anthropic and Meta are, in effect, asking users to hand over a meaningful amount of oversight over parts of their digital life: inboxes, calendars, payroll data, contracts, and in Muse’s case, browser sessions and form submissions. Anthropic’s bet is that most people will trust an agent more if it visibly stops and asks before doing anything consequential. Meta’s bet is that most people will trust an agent more if it simply gets things done without interruption, as long as the end result matches what they asked for. Neither company has published independent data on how often either approach actually goes wrong, which means the real test of Claude Duet versus Muse Wonder will not be feature lists, it will be how each one performs across millions of ordinary, messy, real-world tasks over the coming months.
For now, Claude Duet remains a US-only, waitlist-gated research preview, while Muse Wonder is already further along in its rollout to developers and consumers alike. That gap gives Meta a head start in raw usage, but it also means Meta’s claims about reaching the “frontier” will be tested at a much larger scale, and much sooner, than anything Anthropic has shipped so far with Claude Duet.
Frequently Asked Questions
What is Claude Duet?
Claude Duet is a proactive AI agent from Anthropic, currently in research preview, that identifies tasks needing a user’s attention, performs work on its own, continues working while the user is away, and asks for approval before taking consequential actions.
Who can access Claude Duet right now?
Access is limited to Claude Pro and Max subscribers located in the United States, and requires joining a waitlist. Team and Enterprise customers are not included in this initial preview.
What AI model powers Claude Duet?
Claude Duet runs on Opus 5.5, Anthropic’s top-tier reasoning model, which has also led independent benchmark comparisons this year, including Terminal-Bench 4.0.
Does Claude Duet cost extra on top of a Claude subscription?
Anthropic has not published a separate price for Claude Duet. It is currently included for eligible Pro and Max subscribers who join the waitlist, with no confirmed additional fee.
How is Claude Duet different from Meta Muse?
Claude Duet is designed to prepare and draft work, then pause for a user’s approval before acting on anything consequential. Meta has described Muse as advancing toward a stated goal more independently, including opening browsers, filling out forms, and negotiating on a user’s behalf.
What platforms support Claude Duet?
Claude Duet is available through standalone iOS and Android apps, as well as through the web at claude.ai.
When did Meta launch Muse Wonder?
Meta introduced Muse Wonder and Muse Wonder Full on October 12, 2026, describing them as its most capable models to date for personal and professional agentic tasks.
Is Claude Duet available outside the United States?
Not yet. The research preview is currently restricted to US-based Claude Pro and Max subscribers who have joined the waitlist, with no confirmed international expansion date.
