Claude Duet Rollout: US Max Users In, Pro Next, No Date for Anyone Else

At 09:08 UTC on 28 October, TestingCatalog — an account that tracks staged feature rollouts — posted the first concrete queue for Claude Duet: US Max subscribers were being let in, US Pro subscribers were next in line, and everyone outside the US was told that broader expansion is planned with no timeline attached, with EU users likely to wait longest. Claude Duet is the proactive agent Anthropic opened as a research preview on 21 October, and it runs on Claude Opus 5.5. The queue is the story. Launch-week coverage settled what Duet does; no English-language report we could find had published who gets it, in what order.

The ordering is not Anthropic's own statement, and that distinction matters more than it usually would. Anthropic's release note for 21 October says only that it "just opened a research preview for users on Pro and Max plans." The Get started with Claude Duet help article, updated this week, adds that eligibility requires a US Claude account and a waitlist invitation, and that invitations "go out in waves." Neither document says Max comes before Pro. The Max-first sequence comes from TestingCatalog's 28 October observation. A Chinese-language write-up dated the same day repeated that order without citing a source of its own, so it is repetition rather than independent confirmation. Treat it as observed practice rather than published policy: a company can reorder a wave without editing a help page, and the invite emails are the only record that will settle it.

What Anthropic has actually committed to

Six facts are first-party and stable enough to plan against, all from Anthropic's product page and help centre:

• Who is eligible — Claude accounts in the United States on the Pro or Max plan. Free, Team and Enterprise accounts are excluded for now.

• How access arrives — a waitlist and invitations sent in waves. Joining the waitlist does not guarantee a date, and Anthropic has not published how many invitations go out per wave.

• What it runs on — Claude Opus 5.5, selected by Duet according to the task. Users cannot choose the model.

• What it costs — nothing extra during the preview. Duet draws on the plan's existing usage limits, including the five-hour session and weekly caps, at the same rate as chats.

• What a new admit receives — one full limit reset, valid until 30 November 2026.

• Where it lives — standalone iOS and Android apps plus claude.ai. It is not inside the main Claude app: memory, skills and connected apps are carried over from the Claude account, and the two run as separate surfaces.

The product page is the document to read if you want Anthropic's own framing, and it is careful about what it promises. Duet checks connected accounts "every few hours" rather than continuously — a throttle Anthropic states openly, because a tighter loop would exhaust plan limits faster. It works in a cloud virtual machine with its own browser, which is why it can keep going after you close the app, and why the browser is the part you cannot watch: the preview does not let users view or take over that session. Even the prohibitions are specific. Approval is required before sending messages, making payments, signing documents or deleting anything, and some actions are refused outright even with approval — trading securities, moving money between accounts, and reaching into someone else's account. A dedicated classifier and safety prompt block the strictest data tier, including credit card numbers and government identification numbers, and Anthropic says Duet never directly sees passwords or sign-in tokens.

Why Max first is the order that makes sense

Anthropic has not explained the sequencing, so any reason offered here is inference. The most plausible one is arithmetic. Duet's background work is billed against the same five-hour and weekly allowances as ordinary chat, and Anthropic warns users outright that they may hit limits "sooner than with chat alone." A Max plan carries a multiple of Pro's allowance — five times on the entry Max tier, twenty on the top one — so it is the cohort with the most headroom to absorb an agent that spends tokens while nobody is typing. Opening there first means the first wave of users is the least likely to spend its first day against a cap. The second plausible factor is cost: every active Duet runs its own cloud virtual machine, and preview capacity is finite. Anthropic has published neither the invite volume nor the VM budget, so this stays a reading of the incentive, not a disclosed plan.

What is not inference is the consequence for Pro. When the wave reaches Pro accounts, those users inherit the same agent with roughly a fifth of the ceiling. Duet is not a free extra on a $20 plan; it is a second consumer of the same budget, and the reset that softens the first month expires on 30 November.

What Pro users inherit when their wave lands

Pro is $20 a month, or $17 a month billed annually. The help centre is unusually direct about what an invite does to that plan:

• Limits — the same five-hour session and weekly limits as chat, charged at the same rate. Background checks count too, so a quiet afternoon of monitoring can leave less room for the work you actually asked for.

• Reset — one full limit reset for newly admitted users, valid until 30 November 2026. After that, the preview runs on the ordinary allowance.

• Model choice — none. Duet picks the effort level itself, so there is no way to trade Claude Opus 5.5 down to a cheaper sibling to stretch a weekly window. On a plan this size, that is the missing lever that would otherwise make the preview affordable.

• Monitoring cadence — every few hours, not continuously. Urgent items between checks wait, which Anthropic lists as a known limitation rather than a bug.

• Surfaces — the agent's own cloud browser is invisible and uncontrollable, Microsoft 365 is not supported, Slack support is limited, Duet cannot see every project or local file, and it makes no phone calls or texts, though approved emails are possible.

The practical version: if your Claude usage already brushes the weekly cap, an invite can make the rest of Claude feel smaller rather than bigger. If it does not, Duet is the cheapest way to find out what an always-on agent is worth to you, because the preview is included in a subscription you are already paying for. The people who should be most careful are the ones already running Claude Code on the same Pro plan, because the plan has one allowance rather than two.

Duet is an app, not an endpoint

This is the part that gets flattened in launch coverage, and it is the part developers need. Claude Duet is not callable. There is no model string, no per-token price, no API key that reaches it. It is a subscription surface — app, waitlist, plan limits — and OrcaRouter does not carry it, because Anthropic does not sell it as an endpoint. If you have been waiting for a Duet API, the wait has no published end.

The layer underneath it is callable, and it is the same model. Claude Opus 5.5 is available on OrcaRouter at $4 per million input tokens and $20 per million output tokens, passed through at the provider's list price with no markup added. On the Artificial Analysis Intelligence Index — the v5.0 rebuild dated 27 October — Claude Opus 5.5 reads 58, second of 52 models, with output speed around 97.9 tokens per second. Those are independent figures, not Anthropic's own marketing numbers, and they describe the model rather than the agent.

What the $20 subscription buys on top of that is the orchestration: persistent memory, inherited skills and connectors, a cloud virtual machine with a browser, the approval gates, the classifier that blocks card and ID numbers, and billing that hides all of it behind one plan limit. None of that is for sale separately. If what you need is the reasoning inside a pipeline you already own, the model is the reachable part; if what you need is the finished consumer agent, only the waitlist reaches it. For teams building something Duet-shaped on top of Claude Opus 5.5, running it behind a fallback chain is the low-risk way to start — a preview-grade agent loop is the last place you want a single provider's bad afternoon to become an outage, and automatic failover keeps the request alive without a code change.

Who is still outside the door

Three groups, and none of them has a date. Team and Enterprise accounts are excluded from the preview entirely, though Anthropic says support is planned. That is a real exclusion for the buyers most likely to want an agent that works across business apps: Team seats run $25 a month for standard and $125 for premium, Enterprise is $20 per seat per month billed annually with a 20-seat minimum, and none of those seats can admit Duet today. For an organisation, the preview is not a purchasing decision yet — it is a thing to evaluate when the plan type appears.

Everyone outside the US is in the same position, with the ordering implied rather than published. Anthropic says expansion is planned; TestingCatalog says there is no timeline and that EU users may wait longer than the rest. A waitlist signup is free and costs nothing but an email, so the rational move is to join it and stop planning around it. The mistake to avoid is designing a workflow that assumes an invite arrives in a particular quarter.

Free accounts are excluded and have no path in at all, which is the one clean line in the whole document.

What nobody has published yet

The confidence gap around Claude Duet is wider than the feature list suggests, and a reader deciding whether to care should know exactly where it sits:

• Task performance — Anthropic publishes no benchmark for Duet, no task-completion rate, no latency figure, and we found no independent evaluation of the agent. The reviews circulating this week, including a 7.5/10 scored review with category sub-scores, are early impressions rather than measurements. Artificial Analysis scores models, not consumer agents, so Claude Opus 5.5's 58 tells you about the engine and nothing about whether Duet notices the right email.

• Post-preview pricing — "included in Pro and Max" is a preview term. Anthropic has not said what Duet costs if it leaves research preview, or whether it will consume limits at the current rate when it does.

• Invitation volume — no number of invites per wave, no acceptance rate, no expected wait. TestingCatalog's 28 October post gives no count, and neither does Anthropic.

• Enterprise and international dates — planned, undated, in both cases.

• Browser control — not shipped in the preview, and not promised for it.

What to watch, in order: the first Pro invite wave, which would confirm the Max-first reading; any Anthropic statement on Team and Enterprise timing; the first independent task-completion evaluation, which is the number that would actually settle whether the agent is good; and post-preview pricing, which decides whether Duet stays a bundled feature or becomes a line item. Until the first of those lands, the honest summary is that the queue moved on 28 October and the product is still unmeasured. If you are a US Max subscriber, open the app and check — the invitation may already be there. If you are on Pro, the line is moving and there is nothing to do but wait. And if you need this capability inside your own product rather than inside Anthropic's app, Claude Opus 5.5 is callable today, on one API alongside 200-plus other models, while Claude Duet is not callable at all.