Claude Duet vs Claude Tag: Where Each One Actually Works

Claude Duet and Claude Tag are the two ends of Anthropic's answer to a complaint most knowledge workers have made at least once: you have to go to Claude instead of Claude coming to you. Duet comes to one person. It briefs you in the morning, takes work you hand it by message, and does the work on its own computer in the cloud. Tag comes to a team. It sits in a Slack channel, reads the thread, and reacts while everyone is still in the conversation.

Read both product pages in the same sitting and the interesting comparison is not a feature count. It is that each one is structurally unable to do the other's job, and that each one's restrictions point at the other one's audience. What follows applies six dimensions to both products in the same order. Where Anthropic publishes nothing for a dimension, the entry says not published rather than filling the hole with something reasonable-sounding.

They fail in opposite directions

Duet's failure mode is collective work. The brief lands with one person. The deck comes back to one person. The approval happens in one person's app, and if that person is on holiday the work waits. Anthropic's own page admits a version of this: it may misread a thread or miscalibrate what is urgent, and it can miss something that arrives between check-ins. Those are failures of one agent maintaining one queue on behalf of one human.

Tag's failure mode is the mirror image. It can only work with what is in the thread it was tagged into, which means the work has to have been written down in a channel first, by people who chose to write it down. A team that resolves things in a hallway, a call, or a private message has nothing for Tag to read. The product is not weaker than Duet on that axis; it is a different bet about where the raw material of work lives.

Six dimensions, applied to both

Where it runs

Claude Duet. A standalone app for iOS and Android. On a computer, Anthropic points you at claude.ai in a browser. The work itself happens on its own computer in the cloud, using the apps you connect — so the surface you touch is a phone or a tab, and the thing doing the work is not your laptop.

Claude Tag. In Slack. The route the page offers is an Add to Slack button, and the availability sentence scopes the product to Slack for Enterprise and Team customers. Beyond that surface, no date and no commitment are published.

Who it is for

Claude Duet. One person on a consumer plan, in the United States. The page says Pro and Max plans in the US, and the FAQ is blunter: it is a research preview on a waitlist for Pro and Max subscribers in the US, and Team and Enterprise plans are not included yet. A team administrator cannot buy it for the team today; an individual on Max can join a waitlist.

Claude Tag. An organisation that already holds an Anthropic enterprise tier. Available for Claude Enterprise and Team customers, in Slack. Not Pro, not Max, not the free tier.

Put those two eligibility rules side by side and the buyers are disjoint. Duet's decision-maker is a person with a subscription. Tag's decision-maker is an administrator with a plan. Almost nobody is both, which means almost nobody is choosing between these products in any practical sense.

What it can reach

Claude Duet. More than 6,000 connectors and plugins, per the product page, plus whatever a Claude account already carries: memory, skills and apps are there from the first message. The page names the inbox and the calendar as the places it looks for what needs you, and it works in the apps you connect from inside its own cloud machine.

Claude Tag. The thread it is in. The page's description is about reading threads and understanding full context, and the only named systems in the accounts Anthropic publishes — Datadog, Linear and GitHub — appear inside a customer's description of triaging bug reports, not in a supported-integrations list. Read them as an illustration of reach, not as a roster.

Who approves an action

Claude Duet. The subscriber does, and the page is unusually direct about it. It asks before sensitive operations like sending messages or paying, and it waits for your ok when it matters. The examples follow the same rule: replies are drafted for you to approve, payroll is prepared for you to review, contracts are got ready to send. One person is the approver, and that person is also the beneficiary.

Claude Tag. Not published. Nothing Anthropic publishes states what a human has to approve before an action lands in a channel, or whether an action can land without one. For a team product that is the dimension that decides how much autonomy you actually granted, and it is the one the page leaves open.

Scheduled versus reactive

Claude Duet. Both, in a fixed shape. The morning brief is the scheduled half; it arrives without being asked for. The rest is reactive to you: tasks you hand it by message, work that continues while you are away, and a message back when it is done or needs a decision. Usage ties the two together — during the preview it shares your plan's limits, its background checks count toward them, and larger tasks like a financial model or a deck need more model turns.

Claude Tag. Reactive. The published description is about reading a thread and reacting in real time, and scheduled behaviour is not published alongside it.

What is published about access

Claude Duet. Plan, country, and the interaction with usage limits. Pro and Max, United States, waitlist, Team and Enterprise excluded, background checks counting against the plan. What is not published about access is any administrative control: Duet is described as a personal app for one subscriber, and the page sets out no control list.

Claude Tag. Plan and surface. Enterprise and Team, in Slack. Price is not published — not a per-seat figure, not an add-on, nothing attached to the availability line. Nor is a list of administrative controls, nor a data-retention statement. For a product sold to organisations those are the three questions a security review asks first, and all three are open.

Push versus in-thread

The two products differ less in capability than in who sees the work. Duet pushes. The brief arrives in one app, the dashboard comes back to the same app, and the team sees the artefact rather than the process that produced it. That is efficient and it is private, and it also means the value of the agent is capped by how good one person is at handing work off and approving it.

Tag is in-thread. The work happens where the conversation is, in front of everyone in the channel. The customer quotes Anthropic publishes argue exactly this, and they argue it more clearly than the product description does: one person describes doing work "out in the open in a public Slack channel" with teammates, and another describes managing what Claude can reach "in one place". Visibility is the feature. It is also the risk, since a shared record is a shared failure.

The practical version of this contrast shows up in what you have to change to adopt either one. Duet asks one person to trust a queue. Tag asks a team to write things down in a channel and to let an agent read them, which is a culture change dressed up as an installation.

The gates, not the features

The feature comparison above is the easy half. The hard half is that both products are gated, and the gates do not overlap.

If you are in the United States on Pro or Max, Duet is a waitlist you can join and Tag is not purchasable by you at all. If you are on Enterprise or Team in Slack, Tag is available and Duet is explicitly not included for your plan. If you are outside the United States, Duet's page publishes no date for you and Tag's page gives no date or commitment for any surface other than Slack. The honest summary is that for most readers of a comparison like this one, neither product is buyable today — and that is a finding about the pair, not a knock on either.

The question that actually decides between them is therefore not which is better. It is which shape your work already has: does it arrive as a message to one person with a decision attached, or does it exist as a conversation in a shared channel with a decision in the middle of it? Answer that first, then check whether the gate for that shape is open to you.

What neither one sells you

Neither product is a model you can call, and neither is something OrcaRouter hosts. We host models, not Anthropic's applications, so there is no plan on our side that gets you a Duet invite or a Claude Tag installation. That boundary is worth saying once, plainly, because both products are easy to confuse with the model underneath them.

Duet does publish what it runs on. "Powered by Opus 5.5" is Anthropic's phrasing, and Claude Opus 5.5 is on our catalogue at $4.00 per million input tokens and $20.00 per million output tokens, with a 1M-token context window and cache reads at $0.20 per million. It sits behind the same OpenAI-compatible endpoint as the rest of the catalogue, one key for 200-plus models, at provider list price with no markup, with automatic failover across providers. What that buys you is the engine and the API. It does not buy the morning brief, the app, or anything that lives in Slack.

Claude Tag's page names no model at all, so there is nothing to route there either, and any article that tells you which model powers it is guessing. That asymmetry is worth noticing on its own: the personal product tells you what it is built on and the team product does not.

Read together, the two pages describe a company splitting the same assistant in two directions — outward to one person's queue, inward to a team's conversation — and publishing far more about the first than the second. Six dimensions are enough to see the split clearly. It is also enough to see that the choice between them is mostly a choice about where you work, made in advance by where Anthropic has decided to sell.