Claude Duet: What the Leak Got Right About Anthropic's Proactive Agent, and What Stripe Link Actually Does

The leak that circulated before launch guessed the shape wrong in one place and right in another. It said Claude Duet would be an agent whose interface was "designed around threaded" conversation, and it flagged a shopping flow running through Stripe's Link wallet. The threading call was correct: A​nthropic's own description is that you message Duet "through typing or voice in one ongoing thread," so a quick reply and a weeklong project start in the same place. The shopping call was directionally right but narrower than the leak implied — Link is a payment rail with an approval step, not a shopping surface. And the piece the leak did not mention is the one that decides whether you can use this at all: Duet runs on Claude Opus 5.5 and draws from the same Pro or Max usage allowance as your ordinary Claude chats, background checks included. A​nthropic opened Claude Duet as a research preview on October 21, 2026, for Pro and Max subscribers in the United States, by waitlist. This page is what is confirmed by A​nthropic's own launch material, what the payment flow actually does with your card, and the arithmetic on what an always-on agent costs when every check is billed against a plan limit. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

What shipped, precisely, and what did not

Duet is a separate app — iOS, Android, and claude.ai in a browser — not a mode inside the main Claude app. A​nthropic says the two will relate differently later and that early-user feedback will shape how; for now they are distinct products. It carries over memory, skills and connected apps from your existing Claude account, so setup is mostly authorization rather than configuration.

The research preview's boundaries are unusually specific, and they are the first thing to check against your own situation:

• Who — Pro and Max subscribers in the United States only, by waitlist. Team and Enterprise plans are excluded for now, with no announced date for either.

• What it costs — nothing separate. Duet is included in Pro ($20 a month, or $17 a month billed annually at $200 up front) and Max (from $100 a month, at 5x or 20x Pro usage), and it spends the same usage allowance as the rest of Claude.

• What it runs on — Claude Opus 5.5, which A​nthropic describes as the choice for judgment-heavy work: deciding what needs you, what can wait, and which tool to reach for.

• Where it runs — a cloud virtual machine with its own browser, so it keeps working with your laptop shut. It cannot yet show you that browser, so you get the result and the approval prompt, not the step-by-step.

• What it knows — more than 6,000 connectors and plugins from the Claude Marketplace: Gmail, Slack, G​oogle Calendar and Drive, Microsoft 365, Notion, QuickBooks, Stripe, Gusto, HubSpot, DocuSign, Canva, Shopify, Ramp, 1Password.

The cadence is the part people skim past. Duet checks connected accounts every few hours, not continuously, and A​nthropic gives the reason directly: continuous checking would consume plan limits faster. That single design choice explains most of the product's personality. It is a colleague who looks in three or four times a day and messages you when something is wrong, not a process watching a socket.

One thread, and the interface bet behind it

The leak's read on threading was sharper than it looked. Most assistant products split work across conversations, projects, and separate task views, which puts the burden of remembering where a thing lives on the person. Duet collapses that: a two-word reply and a multi-day build both occupy the same continuous thread, and the agent is expected to hold the difference in context rather than in your head.

That is a genuine UX bet, and it has a failure mode A​nthropic names itself — Duet "may misinterpret threads" or misjudge urgency. With one thread and no visible browser, a misread is harder for the user to diagnose than it would be in a chat product where you can scroll back through a discrete task. A morning brief that quietly missed something looks identical to a morning where nothing happened.

The Stripe Link flow: approvals, virtual cards, and a short list of refusals

This is the part the leak was reaching for, and the confirmed detail is more interesting than "agentic shopping." A​nthropic describes the sequence in four steps, and every one of them is a control rather than a convenience:

• Duet confirms the item with you before it buys. The agent proposes a specific purchase; you agree to that purchase.

• Link, Stripe's digital wallet, then asks you to approve the spend separately, in Link's own approval step rather than inside Duet's conversation.

• Payment runs on a virtual card, so the merchant never receives your real card number.

• Shopify's Shop Pay is supported as a second rail at online stores, and sign-in through credentials saved in 1Password is described as coming soon.

Two approvals for one purchase is not redundancy for its own sake. It separates "this is the right thing to buy" from "this is an acceptable amount to spend," which are different decisions and, in a business context, often different people's decisions. The virtual card is the piece with the longest tail: a card number that lives inside a wallet and is issued per merchant or per spend is a card number that can be revoked without reissuing your corporate card.

What the flow is not: a shopping agent that browses retailers on your behalf and checks out. Business purchases here sit inside connector workflows — the invoice, the supply order, the store — which is consistent with how A​nthropic frames the whole product, around work rather than errands. A​nthropic says plainly that Duet was built primarily around work and is weaker at personal tasks like booking tickets.

Some actions stay closed even when you would approve them. A​nthropic prohibits trading securities and moving money between accounts outright — the approval prompt never appears, because there is nothing to approve. The same document describes a dedicated safety classifier and system prompt that block Duet from sending the strictest tier of information, credit card numbers and government identification, while softer categories such as health or personal finance pass only on explicit instruction. Credential handling follows A​nthropic's Claude Managed Agents design, which means Duet does not directly see your passwords or sign-in tokens.

What a background check actually costs, on published rates

Because Duet consumes plan usage, the honest question is not "is it free with Pro" but "how many checks does Pro buy." A​nthropic does not publish a per-check token count, so the only way to answer is arithmetic on published prices with a stated assumption. Claude Opus 5.5 lists at $4.00 per million input tokens and $20.00 per million output tokens, with cache reads at $0.20 per million — A​nthropic's own published rates.

Take a modest check: 40,000 input tokens of email, calendar and document context, 1,500 output tokens of summary and proposed actions. At list price that is $0.16 of input and $0.03 of output, so about $0.19 per check. Six checks a day is roughly $1.14, or about $34 a month. Heavier checks — a day of unread mail plus two documents — push the input side up faster than the output side, because input dominates the bill at a 5:1 price ratio.

Set that against the plan. Pro is $20 a month and Max starts at $100. If background checks alone can approach $34 a month at list rates on a light assumption, then "included in Pro" is really "metered against a shared allowance that a chat-heavy user also draws on," and the person who hits a limit first is the one running the agent hardest. That is not a criticism of the pricing so much as the thing to plan for: Duet's cost is not its subscription, it is the chats you no longer have room for.

A​nthropic's decision to check every few hours rather than continuously now reads as a product constraint with a price attached, and its advice that you can tell Duet what you want to hear about and how often is a spending control as much as a relevance control.

Who can use it today, and what to do if you cannot

US Pro and Max subscribers on the waitlist, in that order. Everyone else — Team and Enterprise seats, users outside the United States, anyone on the free plan — is waiting on an announcement A​nthropic has not made. The waitlist asks what you would delegate first, which is a reasonable signal of how the team is sequencing access.

There is no API for Duet. That matters more than it sounds: the product is a consumer app, and the agent loop that makes it work — a persistent machine, a schedule, a set of authorized connectors, an approval gate — is not exposed for you to call. If you want that pattern in something you operate, you build it, and the model underneath it is the part you can buy today.

Building the same loop yourself, and where routing fits

A proactive agent is a loop with four parts: a schedule, a context-gathering step, a judgment step, and an approval gate before anything irreversible. Only the third needs a frontier model on every pass. The schedule fires cheaply, the context step is mostly retrieval, and most checks end in "nothing needs you" — which is exactly the kind of decision a smaller model handles without costing Opus money.

That is where the cost structure above becomes a design choice rather than a bill. Route the frequent low-stakes checks to a cheaper tier and escalate to Claude Opus 5.5 only when a check turns up something that needs judgment, and the same six-checks-a-day pattern stops being a $34 line. Claude Opus 5.5 is on the OrcaRouter catalogue as anthropic/claude-opus-5.5, listed at $4.00 per million input and $20.00 per million output with cache reads at $0.20 per million — we pass the provider's list price through with no markup, so an A​nthropic price change lands here the same day rather than at the end of a billing cycle. Claude Sonnet 5.5 and Claude Haiku 5.5 sit on the same key for the cheap passes, which means the escalation policy is a routing rule rather than a second contract and a second SDK.

The other half of the pattern is the approval gate, and Duet's version is worth copying literally rather than approximating. Separate the "is this the right action" confirmation from the "is this an acceptable amount" approval, keep the payment credential out of the agent's context entirely, and make the refusal list explicit — some actions should not have an approval path at all. A​nthropic's own write-up is the best available specification of that design, and it is worth reading for the safety architecture even if you never get off the waitlist.

What is still unverified

Three things will decide whether Duet is a product or a demo, and none of them can be answered from launch material. The first is judgment quality at scale: whether a few-times-a-day check cadence catches the things that matter without producing a stream of low-value pings. The second is whether sharing usage limits with chat survives contact with real users — the earliest public reaction to the announcement centred on exactly this, alongside lock-in to one model family and unease about handing an agent access to a whole working life. The third is the connector surface: 6,000 connectors is a marketplace count, not a promise that each one supports the actions a given workflow needs, and A​nthropic's examples — invoices, payroll prep, contract drafts, store setup — are illustrations rather than a supported-action list.

The one genuinely new thing here is not the agent. Persistent agents with a machine and a schedule have been shipping for a year. It is that a frontier lab put the payment rail, the approval gate and the refusal list in the launch document alongside the feature list, which is the first time the commerce side of an agent has been specified this concretely by the company that built it. Whether that becomes the template everyone copies is the thing to watch over the next month — and the thing to design around now, whether or not you get an invite.