Deep Dive — Claude Duet runs on your Pro quota, not a new bill

Anthropic opened Claude Duet to a US waitlist on Thursday: a standalone agent app for iOS, Android and the web, built on Opus 5.5, that works on its own cloud computer and messages you when something needs a decision. There is no new subscription to buy. Duet ships inside Claude Pro and Max, which is either the best deal in agentic AI or a quiet way of charging you twice — depending on how much of your plan you already use.

That is the whole argument, and both labs made a pricing decision this month that reveals how they see it. OpenAI put its equivalent agent above the cheapest paid tier. Anthropic put its agent on the same meter as your own work. Neither company published a per-agent price, because in both cases the agent's cost is a subtraction from something else you already bought.

What Duet actually does

The mechanics are straightforward once you strip the launch copy. Duet has a virtual machine with its own browser, so it does not need your laptop open, and you hand it work by message or voice in one continuous thread. Each morning it sends a brief on what is slipping across your connected apps. In between, it takes tasks — build a dashboard, prepare a forecast, chase unpaid invoices — and comes back with a finished document, sheet or deck. It carries over the memory, skills and connectors from your existing Claude account, and it can reach the more than 6,000 connectors and plugins in the Claude Marketplace, which is how it gets to QuickBooks, Gusto, Stripe, DocuSign, HubSpot, Shopify and Canva, per Anthropic's launch post.

The permission model is more interesting than the feature list. Before it sends a message, makes a payment, signs a document or deletes anything, Duet shows you what it is about to do and waits. A few things stay off-limits even with your approval: trading securities and moving money between accounts. Payments run through Stripe's Link wallet, which issues a virtual card so your real number is never shared, or Shop Pay at Shopify stores. Credentials are handled the way Anthropic handles them for its managed agents, so the model never sees your passwords or sign-in tokens, and a dedicated classifier plus safety system prompt is meant to stop card numbers and government IDs leaving in a message.

Anthropic is also unusually direct about what is missing. Duet cannot show you its browser yet, it lives in its own app separate from Claude, and personal errands like booking tickets are weaker than the business work it was built around. We covered the launch itself in Claude Duet: Anthropic's always-on agent runs on Opus 5.5; what follows is the part the announcement did not lead with.

The meter is the product

Here is the sentence in Anthropic's own post that should shape how you read everything else. Duet "checks your connected accounts every few hours, because checking nonstop would use up your plan's limits much faster."

That is a product decision made by an accounting constraint. An agent that polls continuously is the thing everyone demoed; an agent that polls every few hours is what you ship when each poll spends from the same allowance as your own conversation. During the research preview, Duet shares your plan's usage limits with the rest of Claude, its background checks count toward those limits, and larger tasks — a financial model, a deck — need more turns and more background work.

The arithmetic on the pricing page makes the squeeze visible. Opus 5.5 lists at $4 per million input tokens and $20 per million output tokens, Sonnet 5.5 at $2 and $10, and Fable 5.1 at $10 and $50, according to Anthropic's price list. One million output tokens of Opus 5.5 at list price costs $20 — exactly one month of Claude Pro. So the subscription is not a discount on tokens; it is a cap on them. A Pro plan is a fixed allowance of model work, and Duet draws from it in the background whether or not you are watching, reading mail and calendar entries turn by turn until the allowance is gone.

Two readers get very different products out of that. If you pay $20 and mostly chat, Duet is close to free — it spends capacity you were not using, and there is nothing else on the market at that price. If you are a heavy Claude Code or Claude Work user, Duet is a competitor inside your own wallet, and the fix is a Max plan starting at $100 a month, which is where the agent stops cannibalising the work you bought the subscription for.

There is precedent for how this fine print lands. Power users already sued Anthropic this year over what the "5x" and "20x" labels on Max actually mean, as we reported in AI power users sue Anthropic over the fine print on Claude's '5x' and '20x' plans. Add an invisible background consumer to a metered plan and the argument gets more expensive, not less. Anthropic has published no usage figures for Duet, no per-agent allowance, and no retention policy for a research preview.

OpenAI priced the same input the other way

OpenAI's Dots, shipped at DevDay on September 29, have the same shape: an agent with its own cloud computer and browser, an identity of its own, more than 4,000 connected apps, and background work it calls proactive research. The pricing is where the two diverge. A first dot is included in ChatGPT Pro and Business Premium, and there is no Plus tier — the entry ticket is the $100 Pro plan, with $200 and $500 tiers above it. Conversations with a dot do not count against your ChatGPT limits; the tasks it starts in Codex or ChatGPT Work do, per OpenAI's launch page. Pro subscribers in the EU, Switzerland and the UK were excluded at launch, and Enterprise workspaces only get Dots if an admin switches them on.

Then OpenAI did the thing that explains its model. From October 30 it halved the $200 plan's included usage in ChatGPT Work and Codex from twenty times the Plus allowance to ten, cut GPT-6 Pro messages in chat from 200 a week to 100, and offered existing subscribers a one-time credit; it also launched a $500 plan whose distinguishing feature is speed, not allowance, The Next Web reported. Raise the floor, trim the middle, sell the top on latency.

Put the two side by side and the philosophies are clean. OpenAI charges at the door: a higher seat price, and talking to your agent is unmetered once you are in. Anthropic charges at the counter: a $20 seat, and every background check is a withdrawal. OpenAI's model is better for the reader who talks to their agent constantly and runs few heavy tasks. Anthropic's is better for the reader who wants an agent without a new bill and does not already saturate their plan. Neither is generous; they are just metering the same scarce input — unattended model turns — at different points in the pipe.

The wall both of them hit is credentials

An agent that cannot log in is a chatbot with a to-do list. Duet can see Gmail, Slack, QuickBooks and Stripe because you connected them, and Anthropic says sign-in with logins saved in 1Password is coming soon with your permission.

That "coming soon" is doing a lot of work, because the credential layer arrived separately this week. 1Password for Agents gives the agent a digital key instead of a password, asks you on your own device which account it wants and why, and fills the credentials itself once you approve — a two-factor prompt rather than a handoff. It supports Muse, ChatGPT Work, Dots and Duet today, with more integrations in progress, SiliconANGLE reported.

The banks said no anyway. Morgan Stanley has started blocking third-party agents from its platforms, and JPMorgan, Bank of America, Citi and Wells Fargo are restricting them, after finding agents inside customer accounts and, in some cases, actions that looked independent of the user's instructions, Axios reported. The bankers' problem is not capability, it is identity: an agent that arrives with the customer's valid login is indistinguishable from the customer, and no bank can currently verify scope or intent. Most are moving toward read-only access — an agent may look at balances, not move money.

Read the three announcements together and the architecture of the next year is visible. Anthropic already refuses money movement even with approval. Payments go through virtual cards that can be capped and killed. The one door nobody has opened is the bank login, because a password manager can tell you which account is being asked for but not whether the agent holding the key will stop at what you approved. The agent economy is being built on rails that can be limited, not on identity that can be proven.

What the skeptics have right

Three objections survive contact with the launch.

Approval fatigue is the first. Duet asks before sending messages, making payments, signing documents and deleting things, and every one of those prompts is a decision a busy person will eventually stop reading. Anthropic's own answer is architectural — approvals run inside your cloud computer, actions are checked by the same auto-mode classifier that guards Claude Code — but the human is still the last line, and a tapped prompt is not consent.

The second is that bundling makes adoption numbers meaningless. Meta's Muse is the only agent in this category with public evidence of demand: more than 10 million users, over a quarter of them buying, average order above $200, per Mark Zuckerberg on Meta's earnings call. Anthropic will be able to report millions of Duet users without anyone paying a dollar more than they already did. That is good distribution and bad measurement.

The third is that no one has published a productivity number. There is no task-completion rate, no time-saved figure, no error rate on a fixed set of jobs — just testimonials from a broker, a restaurant owner and a brand strategist. Compare the diagnostics paper Microsoft published this week, which at least put 304 hard cases and a percentage behind its claim. The always-on agent is being sold on the feeling of the demo.

The counter-case is not weak, though. Distribution is the hardest problem in agents, and Duet has the cheapest on-ramp anyone has shipped: an existing $20 subscription, your existing memory and connectors, no procurement. The quota is the honest constraint. A reader who understands that a background poll costs them a slice of their afternoon's Claude usage can decide whether the trade is worth it, which is more than OpenAI's $100 gate lets them do.

What to watch

Whether Duet leaves the US waitlist and reaches Team and Enterprise, where the quota question gets financial rather than personal. Whether Anthropic publishes an allowance specific to the agent instead of sharing the plan's pool — a separate meter would be the clearest signal that it intends to sell agent capacity rather than bundle it. Whether $20 Pro survives an always-on agent at all, or Pro becomes the tier that gets the agent and Max becomes the tier where the agent works.

And the one that will decide the category: the first Duet incident involving a real account, a real invoice or a real email going to the wrong person. Meta's Muse supplied the cautionary tale in September. Every lab since has shipped a permission model designed to survive that story, and none has been tested by it yet.

If your agent is only as good as the quota you are willing to give it, would you rather pay $20 and share, or $100 and not think about it? Write to us via our contact page.