# Claude Duet: 4 Approval Gates, 2 Plans [2026] > Source: > Published: 2026-10-30 19:30:43+00:00 Anthropic has put [Claude Duet](https://shattered.io/claude-duet-professionals-not-consumers-3-apps-2-plans-2026/) into research preview, a new app where Claude works as a [proactive agent](https://claude.com/ko/resources/articles/claude-duet?li_fat_id=f980db50-9530-46c9-9ef0-3fad8d2665d9) that surfaces what needs your attention, completes tasks you ask for, keeps working while you are away, and checks with you before it acts on your behalf. The launch is the newest move in a crowded race, and it matters because Anthropic is aiming Duet at working professionals instead of general consumers. Coverage this week, including Tatler Asia (dated October 29, 2026), Herald Economy, news.lavx.hu and Education Times, describes Duet as a digital chief of staff. That label is the press shorthand for an agent that watches your work in the background, not a formal Anthropic product tier. This analysis sorts what Anthropic has confirmed from what outlets have reported, because the two lists are not the same length. The timing is not an accident. Rivals have spent months wiring their assistants into email, calendars and payment tools, and we tracked one version of that push when [Google opened Gemini and Search to 50+ apps](https://shattered.io/google-connected-apps-platform-gemini-search-50-apps-2026/). Duet is Anthropic’s answer, and it arrives wrapped in more caution than most of its competitors have shown. 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 Claude Duet Is, and What Anthropic Has Confirmed Start with the facts that survive scrutiny. Duet is an Anthropic product, it is in research preview, and Anthropic describes it as an app where Claude behaves as a proactive agent. Four capabilities sit at the center of the pitch. Duet surfaces items that need attention, completes requested tasks, continues work while the user is away, and checks with the user before acting for them. Those four verbs describe a different relationship than a chatbot offers. A chatbot waits for a prompt, answers, and goes quiet. Duet is built to start the conversation, hold a to-do list across hours, and come back with a question or a finished piece of work. Anthropic’s help center frames the interaction as messaging Claude the way you would message a person, and says you can type or speak your requests. Duet also does its work on a cloud-based computer, which is a quiet but important design choice. The agent does not borrow your laptop while you sleep. It runs somewhere else, which is how it can keep going after you close the app. That architecture explains both the appeal (always on) and the risk (a remote machine acting with your connected accounts). Anthropic built the product for professionals and small-business owners, according to its announcement. The company’s own wording points at people with more on their plates than they can track, which is a recognizable description of a founder, a consultant or a team lead without an assistant. Notice what is missing from that audience: large enterprise IT departments. We come back to that gap below. ## How the Proactive Agent Works The word proactive carries most of the product’s weight, so it deserves a plain definition. A proactive agent does not need you to open it first. It looks at the apps you have connected, decides something deserves your eyes, and tells you. It can also take on a delegated task, work through it on its own, and report back when it finishes or hits a decision that belongs to you. ### The cloud computer behind the agent Because Duet uses a cloud-based computer, its work persists independently of your phone’s battery or your laptop’s lid. Tatler Asia describes the setup as an independent cloud workspace that splits tasks into organized project channels, so client questions, travel bookings and research do not pile into one endless thread. Anthropic’s confirmed language is narrower, and the channel detail comes from Tatler’s reporting rather than the official sources I retrieved. The practical upside is continuity. Hand Duet a task before a flight and the work can move forward while you are in the air. The practical cost is trust, since you are delegating access to accounts you may not be able to supervise in real time. Anthropic’s answer to that cost is the confirmation design covered in the next section. ### Connected apps and the tools it touches Yahoo Tech reports that Anthropic says Duet can connect to more than 6,000 apps and services, with examples ranging from Canva and Slack to QuickBooks, Stripe and LinkedIn. That figure comes from press coverage of the announcement and does not appear on the verified fact list I worked from, so treat it as reported. Tatler lists a similar set of connection targets and mentions the Model Context Protocol as the plumbing. Whatever the exact count, the principle is the same one every agent vendor now follows. The assistant is only as useful as the systems it can read and write. Our explainer on [zero trust for AI agents and the three tiers after login](https://shattered.io/zero-trust-for-ai-agents-three-tiers-after-login-toledano-2026/) covers why that access, once granted, becomes the real security question. ## The Four Approval Gates: Message, Payment, Signature, Delete The most specific safety promise in the launch is also the easiest to remember. Duet is designed to ask for confirmation before it sends a message, makes a payment, signs a document, or deletes anything. Yahoo Tech cites Anthropic’s blog as saying the agent shows you exactly what it is about to do and waits for approval. That is four gates, and each maps to an action that is hard or impossible to undo. | Action | Duet behavior | Why it matters | Source | |---|---|---|---| | Send a message | Asks for confirmation first | An outgoing email or text cannot be recalled once delivered | Anthropic, via Yahoo Tech | | Make a payment | Asks for confirmation first | Money movement is the costliest error an agent can make | Anthropic, via Yahoo Tech | | Sign a document | Asks for confirmation first | A signature can create a legal commitment | Anthropic, via Yahoo Tech | | Delete anything | Asks for confirmation first | Deleted files and records may not come back | Anthropic, via Yahoo Tech | | Trade securities | Off limits even with approval, per Yahoo Tech | Removes the highest-stakes financial misuse case | Yahoo Tech, TechTarget | | Move money between accounts | Off limits even with approval, per Yahoo Tech | Blocks the most direct path to fraud or error | Yahoo Tech, TechTarget | The last two rows go beyond the confirmation list and come from press reports of Anthropic’s statements, so I label them accordingly. TechTarget adds that Anthropic uses a dedicated classifier and safety-system prompts that refuse to share sensitive items such as credit card numbers and government IDs, and that users can direct Duet to make business purchases through Stripe’s digital wallet. Gates only help if people read them. The usual failure mode for approval prompts is fatigue, where users tap accept without looking. A well-designed preview screen that shows the exact message, amount or file in question gives Anthropic a better shot at avoiding that trap than a vague confirmation dialog would. ## Who Can Use It Now: Pro, Max and a US Waitlist Access is narrow on purpose. Duet is available through a waitlist to users on Pro and Max plans in the United States. Nobody on a free plan is in the first wave, and nobody outside the US is either. Yahoo Tech notes the waitlist is open to Pro and Max subscribers, which matches the official framing. On platforms, Duet is described as a standalone app for iOS and Android, and it also runs on the web through Claude.ai. TechTarget reports the same lineup. A standalone mobile app is a meaningful choice because approval prompts need to reach you where you are, and a phone is where most people are when an agent needs a decision. Pricing is the open question. The sources I retrieved do not establish an official price for Duet itself. Tatler Asia reports that the preview sits with subscribers paying upward of US$20 a month for Pro and Max, which reflects the plan tiers rather than a separate Duet fee. Anthropic has not published a standalone price, and anyone quoting one is working from inference. The Pro and Max gate also serves a quiet product purpose. A paid, waitlisted, US-only preview lets Anthropic watch how an always-on agent behaves with a limited population before it scales the infrastructure and the legal exposure. Compare that with the splashier consumer launches that rivals have staged, and the contrast is deliberate. ## Built for Professionals and Small-Business Owners Yahoo Tech points out that Anthropic is billing Duet specifically as a work tool, unlike some rival assistants that lean into consumer appeal. In its blog post, Anthropic says professionals of all kinds juggle many decisions a day, and it pitches Duet as help with prioritizing time and energy so there is room for ambitious projects. That focus shapes the feature set. Yahoo’s report says Duet can take on tasks like drafting job descriptions or assembling reports, which are the unglamorous chores that eat a small-business owner’s week. The value proposition is not a clever essay or a code snippet. It is the reclaimed hour in a packed afternoon. Yahoo also observes a stylistic split. Dots and Muse drew attention for a colorful, cuddly look that critics read as a way to lower the guard of users handing over personal data, while Duet so far lives as a plain assistant you talk to inside an app, with no cute avatar. Whether that restraint helps or hurts adoption is an open bet, but it fits a brand that sells on seriousness. The small-business angle also makes commercial sense. Solo operators buy tools with their own credit cards and decide in a day. Enterprises take months. If Anthropic wants early revenue and fast feedback from an always-on product, the person running a five-person shop is a better first customer than a Fortune 500 security committee. ## What Remains Unconfirmed Good news analysis tells you where the ice gets thin. Several details repeated in headlines and explainers are not confirmed by the official Anthropic sources I retrieved. I list them here so you can weigh them yourself. | Claim | Where it appears | Status | |---|---|---| | Duet is powered by Claude Opus 5.5 | TechTarget, Tatler Asia | Reported by outlets, not confirmed by retrieved Anthropic sources | | Rivals are named OpenAI’s Dots, Meta’s Muse and Google Gemini’s Spark | Tatler Asia, Yahoo Tech | Reported by outlets, not named in retrieved Anthropic sources | | Duet runs background sweeps every three hours | Tatler Asia | Unconfirmed | | Duet sends a morning briefing | Tatler Asia | Unconfirmed | | Duet directly pings users when something is urgent | Tatler Asia | Unconfirmed | | Official price for Duet | No retrieved source | Not established | None of these claims is absurd. A model like Opus 5.5 is a plausible engine, and a morning summary is the natural feature for a chief-of-staff product. But plausible is not confirmed. When Anthropic or a named executive states these details on the record, the table above changes. Until then, readers should treat the Opus 5.5 link, the three-hour cadence and the briefing as press descriptions. One more gap is worth flagging. I found no on-record quotation from a named Anthropic executive about Duet in the retrieved material. The company speaks through its blog and help pages, so every attributed statement in this piece traces to a corporate document or a press report of one. ## The Competitive Field: Dots, Muse and Gemini Spark Duet does not launch into an empty market. Yahoo Tech calls it the latest entry in a run of personalized agents from major AI companies, and both Yahoo and Tatler name OpenAI’s Dots, Meta’s Muse and Google Gemini’s Spark as the competition. The reports disagree on some dates, which tells you how fast the category is moving and how loosely the names are being used. | Agent | Maker | Reported timing | Reporting outlet | |---|---|---|---| | Claude Duet | Anthropic | Research preview, October 2026 | TechTarget, Yahoo Tech, Tatler Asia | | Dots | OpenAI | Introduced in September | TechTarget, Yahoo Tech | | Muse | Meta | Launched in September, described as an instant success | Yahoo Tech, TechTarget | | Gemini Spark | | First launched in May at Google I/O, per Yahoo Tech | Yahoo Tech | | Gemini Agent | | Released October 8, per TechTarget | TechTarget | | Claude Cowork | Anthropic | Released in January, mainly non-proactive | TechTarget | Note the Google rows. Yahoo says Gemini Spark debuted in May at Google I/O, while TechTarget says Google released Gemini Agent on October 8. Those may be two products, two stages of one product, or two labels for the same rollout. I do not have the evidence to decide, so I list both with their sources. ### How reviewers are sizing them up Hands-on comparisons are arriving fast. Wired published a [test of every major AI agent over a single day](https://www.wired.com/story/i-tested-every-major-ai-agent-for-a-day-heres-the-best-one/), a format that suits a category where the product pitch and the daily reality can diverge. I am not relaying its verdict here, since the pitch for each tool deserves a direct read rather than a summary from me. Meta’s side of the race also has a money story. We covered the company’s financing push when [Nvidia put $3B into Meta’s $95B stock raise](https://shattered.io/nvidia-3-billion-meta-equity-offering-muse-vera-rubin-2026/), and the same compute appetite that funds Muse sits behind every agent on this list. On OpenAI’s side, the latest model pricing and product tiers are in our rundown of [GPT-6.1 Terra in ChatGPT Work and Codex](https://shattered.io/gpt-6-1-terra-chatgpt-work-codex-api-pricing-2026/). ## Why Anthropic Arrives Late and Cautious TechTarget lays out the sequence plainly: OpenAI introduced Dots in September, Google released Gemini Agent on October 8, and all three followed what the outlet calls the instant success of Meta Muse. Against that calendar, Duet looks like a follower. Forrester analyst William McKeon-White told TechTarget that Duet appears a bit late, though he credited Anthropic for taking a more cautious approach than its competitors. McKeon-White’s read, as TechTarget reports it, is that nothing about Duet is especially unique apart from the restraint with which it is being introduced. He added that this is the better approach for agents with this level of autonomy, since the internet is full of examples of unintended and rogue behavior. In other words, the analyst treats the slow rollout as a feature. Yahoo Tech sets the same point in a broader frame. The launch comes while AI models face increased scrutiny over security, prompted by a number of rogue AI incidents reported in recent months. A vendor that rolls out an always-on agent into that climate has strong reasons to ship gated, waitlisted and US-only. There is a competitive logic too. Being second or third lets you study the failure modes of the first movers. If Dots or Muse stumble in public, Anthropic can point to its approval gates. If they do not, Anthropic has still bought itself a few weeks of observation. Caution is a strategy here, and it costs little as long as the category does not lock in. ## Enterprise Readiness: The Gap Analysts See TechTarget’s headline says it directly: Anthropic rolls out Claude Duet, but enterprises are not ready. Paul Baier, CEO and analyst at GAI Insights, told the outlet that business-to-business customers will not roll out Duet for months, maybe years, maybe never. He said his firm has no large corporate customers excited about it or planning to use it in the next six months. Baier also pointed to Claude Cowork, the mainly non-proactive agent Anthropic released in January, noting that many enterprises have not even switched it on. His reasoning is that letting non-technical staff chat with AI is low risk, while a personal agent that automates processes touching confidential data and IT systems raises the stakes for every security and governance group. He went further, according to TechTarget, saying no company over 1,000 people is clamoring to roll out Dots, Muse or Duet in the fourth quarter. That is one analyst’s view, and it comes from an advisory firm talking to its own client base. It is still a useful counterweight to launch-week enthusiasm. The takeaway for IT leaders is practical. Duet’s design assumes one person delegating their own accounts. Enterprise rollouts need admin controls, audit trails, data-residency answers and procurement paperwork, none of which the launch material covers. Until Anthropic fills those gaps, the first corporate use will likely arrive through individual employees on personal plans, which is exactly the shadow-IT pattern security teams dislike. ## Historical Context: From Chatbot to Cowork to Duet Anthropic’s path to Duet runs through a few distinct steps. Claude began as a conversational assistant that answered when asked. Claude Code then pushed the model into software work, where permission handling became a daily design problem. One recent example is our coverage of [Claude Code 2.1.311 letting users skip rm checks](https://shattered.io/claude-code-2-1-311-even-more-dangerously-skip-permissions-2026/), a reminder that the tension between speed and safeguards is old news for this company. Claude Cowork, released in January per TechTarget, is described as a mainly non-proactive agent. Duet is the next rung. The agent now initiates, persists and acts across connected apps, which is a much larger surface than a coding terminal. The same approval instinct appears in both products, but the stakes rise when the thing being approved is a payment or a signature. The wider industry went through a similar arc. First came chatbots, then copilots inside single apps, then tool-using assistants, and now persistent agents with their own compute. Each step handed the software a little more initiative and asked the user for a little more trust. Duet, Dots, Muse and Gemini’s agents are all points on that same curve, and the argument is shifting from can it do the task to who is accountable when it does. ## Security and Governance: Where an Always-On Agent Gets Risky An agent that reads your inbox, calendar and accounting tools has the same view of your working life that a human assistant would. It also never sleeps and can be manipulated by text it reads. That second point is the heart of [prompt-injection](https://shattered.io/prompt-injection-test-lab-llm-harness-2026/) risk. A malicious email can contain instructions aimed at the agent, not the human, and the agent has to decide which voice to trust. Approval gates blunt that threat because a hijacked agent still needs a human to tap accept before a payment or a message goes out. They do not eliminate it. The agent can still read sensitive data, summarize it somewhere unexpected, or prepare a convincing but wrong draft. TechTarget’s mention of a classifier and safety-system prompts against leaking card numbers and government IDs suggests Anthropic is layering defenses, not betting on one. For small businesses piloting Duet, a short checklist helps. Connect only the apps the agent genuinely needs. Keep payment methods limited and review each approval screen instead of tapping through. Separate personal and business accounts where possible. Revisit the connections monthly and remove anything unused. None of this is exotic, and all of it shrinks the blast radius if something goes wrong. Anthropic also talks about safety in public. Our report on [Anthropic urging courts to block risky AI models](https://shattered.io/anthropic-newsom-letter-courts-block-risky-ai-models-oct-15-2026/) shows the company arguing for guardrails in public while it ships products that test them in practice. Duet will be read through that lens. ## Market Impact: Subscriptions, Stickiness and the IPO Story Gating Duet to Pro and Max plans turns it into a reason to pay, or to pay more. Agent features are among the stickiest in software because they accumulate context. Once Duet knows your clients, your files and your habits, switching to a rival means rebuilding that memory. That is a retention play, and retention is what public-market investors reward in subscription businesses. That matters because Anthropic is a company under a financial microscope. Our breakdown of the [Anthropic S-1 and its five big questions](https://shattered.io/anthropic-s-1-80-7b-loss-five-big-questions-ipo-2026/) lays out why growth quality, not just growth, is the issue. A proactive agent that converts casual chat users into daily workflow users is the kind of product that supports a durable-revenue argument. The risk runs the other way as well. Always-on agents consume compute while the user is away. If Duet proves popular, heavy background use could strain margins. Anthropic has said nothing public about how it will price that load, and I will not guess at numbers. For rivals, the lesson is that agent launches are now a monthly event, not a yearly one. OpenAI, Meta and Google each moved within weeks of one another, and Anthropic followed. Expect feature parity to arrive quickly, which pushes the real contest toward trust, reliability and integration depth. The vendor that gets small-business owners to hand over the keys without incident wins the category. ## Five Predictions for the Duet Era These are my judgments based on the reporting above, not confirmed plans from any company. Treat them as informed expectations that could change. - **The waitlist widens beyond the US.** A US-only gate for Pro and Max users reads as a first phase. Expect Anthropic to extend access to more regions and possibly more plans once it has watched real usage. - **Enterprise controls follow, but slowly.** Analysts quoted by TechTarget see little corporate appetite this quarter. Anthropic will likely add admin and audit features before it markets Duet to large companies. - **The off-limits list grows.** If a public incident involves any agent, vendors will add new blocked categories beyond trading and account transfers. Approval gates tend to multiply after the first headline failure. - **Pricing becomes the next battleground.** With no official Duet price established, the first rival to undercut on a work-agent tier will force a response.[Usage limits](https://shattered.io/chatgpt-vs-claude-pricing-usage-limits-pro-500-opus-5-5-2026/) will matter as much as sticker price. - **Reviewers shift from features to reliability.** Early coverage lists what agents can do. By winter, hands-on tests like the one Wired ran will focus on how often agents get routine chores right without supervision. Any of these could miss. The category is only weeks old, and one security incident or one viral failure could reorder the list. What I would bet on is direction: more autonomy, more gates, and more argument over who bears the cost of mistakes. ## What to Watch Next Four signals will tell you whether Duet is a lasting product or a launch-week curiosity. First, watch whether Anthropic confirms the model behind it, which would settle the Opus 5.5 question that outlets currently report without official backing. Second, watch for a published price or a clear statement that Duet stays bundled with Pro and Max. Third, track how the preview handles its first public mistake. Every agent will make one, and the response matters more than the error. A fast disclosure and a visible fix would reinforce the cautious brand. Silence would undercut it. Fourth, watch the competitor calendar. If OpenAI, Meta or Google push a new work-focused tier in November, Anthropic’s late-but-careful positioning faces its first real test. For everyday readers, the sensible move is to join the waitlist if you are on Pro or Max and fit the professional or small-business profile, start with low-stakes tasks, and keep the approval prompts switched on. For IT and security leaders, the sensible move is to write a policy now, because employees will try this on their own before any official rollout arrives. ## Frequently Asked Questions **What is Claude Duet?** Claude Duet is a new Anthropic app in research preview where Claude works as a proactive agent. It surfaces items that need your attention, completes tasks you request, continues working while you are away, and checks with you before acting on your behalf. **Who can use Claude Duet right now?** Access runs through a waitlist for users on Pro and Max plans in the United States. It is described as a standalone app for iOS and Android and is also reachable on the web through Claude.ai. **How much does Claude Duet cost?** The sources I retrieved do not establish an official price for Duet. Tatler Asia ties the preview to Pro and Max subscriptions, which start around US$20 a month, but that reflects the plans, not a separate Duet fee. **Will Claude Duet act without my permission?** It is designed to ask for confirmation before it sends a message, makes a payment, signs a document, or deletes anything. Yahoo Tech also reports that trading securities and moving money between accounts are off limits even with approval. **Is Claude Duet powered by Claude Opus 5.5?** TechTarget and Tatler Asia both report that it is, but the official Anthropic sources I retrieved do not confirm the model. Treat the Opus 5.5 link as reported, not verified. **How does Claude Duet compare with OpenAI’s Dots, Meta’s Muse and Google’s Gemini Spark?** Yahoo Tech and Tatler Asia name those three as rivals, though Anthropic’s own material does not. Duet’s reported differences are its work focus, its plain interface without a cartoon avatar, and its cautious, gated rollout. **Does Claude Duet run background sweeps every three hours and send morning briefings?** Tatler Asia reports both behaviors, including direct pings to the user. Neither is confirmed by the official Anthropic sources I retrieved, so they remain unverified descriptions. **Is Claude Duet ready for large companies?** Analysts quoted by TechTarget say no. Paul Baier of GAI Insights expects corporate rollouts to take months or longer, and Forrester’s William McKeon-White praised the cautious launch while calling the product itself not especially unique.