Claude Duet Pricing and Usage Limits: The Number Anthropic Left Out

The number that would settle most of the questions about Claude Duet is the one Anthropic has not published: the token cost of a single background check. Everything else about the economics is either stated plainly on the product pages or can be reasoned about from the model that powers it, Claude Opus 5.5, whose per-token prices are public. But the per-check figure is absent, and its absence is not a rounding error in the documentation. It is the difference between being able to tell a Pro subscriber what a month of Duet will feel like and having to describe the shape of the exposure instead. So this piece does the second thing: it gathers every figure Anthropic actually put its name to, states clearly what is missing, and then shows the piece of arithmetic that does not require the missing number. If you searched for Claude Duet pricing because you wanted a per-task dollar figure, the honest answer arrives in the first few paragraphs, and it is that no such figure exists yet. What follows is what can be said without inventing it.

The number Anthropic did not publish

Anthropic's Duet announcement is unusually specific about the shape of the usage model and unusually silent about its size. It says Duet shares your plan's usage limits with the rest of Claude during the research preview, that its background checks count toward those limits, and that larger tasks such as building a financial model or a deck require more model turns and background work. Those three statements are enough to know that Duet is metered against your subscription rather than sold as a separate add-on, and they are not enough to know what a given week of use will cost you in allowance.

Four specific things are missing, and it is worth naming them so you can stop looking for them. The token cost of a single check is not published. Whether the polling interval is fixed at a constant or adapts to how much activity your connected accounts are generating is not published. What a Pro-tier user can realistically run in a day, as opposed to what a Max 20x user can, is not published. And whether background checks draw from the same pool as interactive messages or from a separate sub-budget is not published either — the announcement says checks count toward the limits, which rules out a wholly separate pool, but does not say whether the split is a sub-budget, a priority ordering, or a flat share of one number.

None of that is a criticism of the documentation. It is a description of a product in research preview, where the operating parameters are the thing being tuned. But it does mean that anyone quoting you a per-task cost for Duet is quoting an estimate dressed as a fact, and it means the useful question is not "how much per task" but "how does this class of product spend allowance at all."

What is published, and where it comes from

The plan prices come from Anthropic's own Cowork product page as it reads in October 2026, and they apply to Duet because Duet is included in Pro and Max rather than priced separately. Pro is 17 dollars a month with the annual subscription discount, which is 200 dollars billed up front, or 20 dollars if billed monthly. Max 5x is 100 dollars a month and Max 20x is 200 dollars a month. Team is 20 dollars per seat per month for teams of two to 150 people. Enterprise is quote-based, and Anthropic has not said when Duet reaches it — the research preview is Pro and Max, in the United States, by waitlist.

Two other published statements matter more than the price list, because they are the ones that govern what a month actually feels like. The first is that Duet runs on the same usage limits as the rest of Claude. The second is Anthropic's own explanation of why Duet polls every few hours rather than continuously: checking nonstop would use up your plan's limits much faster. That sentence is the entire pricing philosophy in one line. The interval is not a technical compromise. It is a budget decision, made on your behalf, published in plain language, and the only lever Anthropic has handed you is an implicit one — how much you connect, and how much you ask it to chase.

There is one more published sentence worth holding onto, and it does not come from the Duet page at all. Anthropic's Cowork page states that Cowork consumes limits faster than Chat, and its FAQ confirms that agentic tasks consume more capacity than regular chat. Two different products, one company, the same finding: the moment software starts coordinating tool calls and taking multiple model turns, it stops resembling a conversation on the billing side. Duet is that same reality on a timer. A chat turn is one exchange you initiated. A Duet sweep is a scheduled sequence of reads, comparisons, and turns that happens whether or not anything was waiting for it.

Every few hours is a budget decision wearing a frequency

Here is the part of the arithmetic that does not need the missing number. "Every few hours" is a range, not a point. Read generously, it could mean four checks a day. Read literally, a check every three hours is eight sweeps a day. Neither reading is published, which is exactly the problem, but both are inside the plain meaning of the phrase, and the band between them is where a Pro user's realistic ceiling lives.

• Four sweeps a day is the forgiving reading. It is what you get if Anthropic means morning, midday, late afternoon, and overnight, and it is roughly the cadence of a product that wants to hand you a brief rather than to monitor anything.

• Six to eight sweeps a day is the literal reading of "every few hours." At three-hour intervals across a waking day you land at six; across a full twenty-four hours you land at eight.

• Whatever you ask for on top is unpriced in every sense. Duet also does the tasks you hand it, and those are the ones the announcement flags as needing more turns and more background work.

That is the shape of the exposure, and it is worth stating plainly what it means: a proactive agent converts wall-clock time into token spend in a way an interactive chat never does. When you close a chat window, the spending stops. When you leave Duet connected, the spending continues on a schedule, and the only published control over that schedule is the interval Anthropic chose. The reason this feels unfamiliar is that consumer software has trained everyone to think of idle time as free. For an agent that polls, idle time is the product, and it is also the bill.

Where the tokens actually go

To reason further you have to know what a sweep consists of, and this part is general to how agentic loops work rather than specific to Duet. A sweep opens with a system prompt and a tool schema — the instructions and the list of actions available — then pulls in whatever the connectors returned, then runs some number of model turns over that material, then produces something: a brief, a draft, a decision to do nothing. The system prompt and tool schema are the same bytes on every sweep. That repetition is the whole reason prompt caching exists as a feature, and it is why the cost structure of a polling loop is dominated by the question of cache hits rather than by the size of the model's output.

The consequence for anyone reasoning about Duet from the outside is direct. If Anthropic keeps the system prompt and tool schema stable across sweeps — and there is every architectural reason to — then the repeated part of each sweep is being read at cache prices rather than full input prices, and the variable part is bounded by how much actually changed in your connected accounts. If they do not keep it stable, every sweep pays full input on the whole instruction block. That single detail, unpublished for Duet, is probably the largest single factor in what a sweep costs, larger than the model choice and larger than the output length.

What the model's own prices tell you

The model underneath is where public numbers exist. Claude Opus 5.5 on OrcaRouter is 4.00 dollars per million input tokens and 20.00 dollars per million output tokens, with a 1M-token context window, and cache reads at 0.20 dollars per million — the provider's list price passed through at 0 percent markup, through one OpenAI-compatible key that reaches more than 200 models. Those five numbers are published, checkable, and they are what make the cache argument concrete rather than theoretical. A cache read costs one twentieth of a fresh input token. If the stable prefix of a sweep is large relative to the changed material, caching is not an optimisation; it is the difference between a sweep that is affordable at a frequent interval and one that is not.

To be precise about what this comparison is and is not: Claude Duet is not billed at API rates. It is a subscription with limits, and your Pro or Max payment is the entire bill. Nobody at Anthropic is sending you a token invoice for a background check. The API prices matter to a different reader — the one who wants the loop Duet implements but cannot buy Duet, because Duet is a consumer app with no published API and no way to re-point it at anything. For that reader the question is not what Duet costs. It is what the equivalent loop costs to build, and there the published token prices and the published cache-read price are the only real inputs, because they are the only ones anyone has published at all.

The three questions to ask instead of a price

If a per-task figure does not exist, what should a buyer or a team lead actually ask? Three things, and all three are answerable from documentation or from your own account's behaviour once you have access.

• How much of my limit does a quiet week consume? Connect nothing meaningful, leave Duet running, and watch. That number is pure polling overhead and it is the floor your month sits on regardless of what you ask for. It is also the number Anthropic has not published, which is why measuring it yourself is the only route.

• Does a task-heavy day crowd out interactive use? The announcement's wording — checks count toward the limits — implies one pool. If you find that a Duet-heavy morning leaves your chat allowance thin by afternoon, you have learned the answer empirically, and you have learned it cheaper than any analyst could tell you.

• What changes the interval? Anthropic has described the interval as a design choice, not a setting. Whether it responds to account activity, to plan tier, or to how much you have already spent this month is unpublished. Watch whether the cadence you observe is steady across a week and across your billing cycle; a constant cadence and an adaptive one look identical for a day and very different over a month.

None of this makes Duet expensive or cheap. It makes it unmeasured, which is a different and more useful thing to know going in. The plain-language reading of Anthropic's own documentation is that the company expects Duet to consume allowance noticeably — that is why the interval is hours and not minutes, and why the announcement pre-emptively explains the choice rather than leaving it to be discovered. A product that polls on a timer and shares your plan's limits is, by construction, a product whose cost is a question of what you leave connected and for how long.

So the practical advice is unglamorous. Treat the first month as instrumentation rather than evaluation. Connect one account rather than six. Watch what a week of polling alone does to your remaining allowance before you hand it a real workload. And if your actual need is a scheduled loop pointed at systems Duet cannot reach, price it on the numbers that are published — 4.00 dollars per million input tokens, 20.00 per million output, 0.20 per million cached reads — and remember that the caching line is the one that decides whether a frequent sweep is a rounding error or the majority of the bill. That is a genuinely different purchase from a subscription, made by a different kind of buyer, and conflating the two is how most comparisons of agent pricing go wrong.