Claude Duet Reads 6,000 Apps but Cannot Move Your Money: Anthropic's Pre-Filing Agent Sends the Toll to Plaid and Visa insight cover
Private CompanyINTU · HUBS · DOCU•18 min read

Claude Duet Reads 6,000 Apps but Cannot Move Your Money: Anthropic's Pre-Filing Agent Sends the Toll to Plaid and Visa

Anthropic's Claude Duet, launched Oct. 21, 2026 as a US-only research preview days before its public S-1 is expected to land, turns a chat window into an always-on agent that reads email, calendars and business ledgers, then pays merchants through Stripe Link or Shop Pay. The asset it assembles is not a better model but a permission graph: more than 6,000 connectors, a live Intuit integration and a Plaid bank rail that ChatGPT and Meta's Muse also rent. Because Anthropic blocks transfers between accounts and securities trading, banks capture nothing while the aggregator and card networks capture the toll, which is why Visa and Nvidia sit on the winning side of this launch and the software an agent merely reads, Intuit and HubSpot, have already been repriced.

Published Oct 22, 2026Updated Oct 22, 2026

Anthropic revenue, Q3 2026

$16.8B

Up from $11.5B in Q2 2026, per Financial Times reporting, Oct. 21, 2026

Connectors available to Claude Duet

6,000+

Claude Marketplace connectors, including Gmail, QuickBooks, Stripe, HubSpot and Shopify

Financial institutions behind AI bank linking

12,000+

Plaid coverage behind ChatGPT's personal-finance preview, May 2026

Intuit enterprise value to revenue

4.4x

At the Oct. 22, 2026 close, down from 11.9x at the Jul. 31, 2025 fiscal year-end

Anthropic revenue, Q3 2026

$16.8B

Up from $11.5B in Q2 2026, per Financial Times reporting, Oct. 21, 2026

Connectors available to Claude Duet

6,000+

Claude Marketplace connectors, including Gmail, QuickBooks, Stripe, HubSpot and Shopify

Financial institutions behind AI bank linking

12,000+

Plaid coverage behind ChatGPT's personal-finance preview, May 2026

Intuit enterprise value to revenue

4.4x

At the Oct. 22, 2026 close, down from 11.9x at the Jul. 31, 2025 fiscal year-end

Claude payers on the $100-a-month Max plan

7.3%

a16z consumer panel, August 2026, versus 1.1% at ChatGPT and 1.3% at Google

The launch, verified

What Anthropic Actually Shipped On Oct. 21

Anthropic announced Claude Duet on Oct. 21, 2026 as a standalone app on iOS, Android and claude.ai that works in the background rather than waiting to be prompted. It runs Claude Opus 5.5 on its own cloud virtual machine with a browser, checks connected apps every few hours, sends a short morning brief and messages the user only when something needs a decision. Access is a waitlist research preview for US Pro and Max subscribers, included at no extra cost, with background work drawing on the same usage allowance as ordinary chat.

  • Reads mail and calendar across connected apps, then assembles a daily brief of what actually needs attention.
  • Reaches more than 6,000 Claude Marketplace connectors, including Gmail, Slack, QuickBooks, Stripe, HubSpot and Shopify.
  • Finds unpaid invoices in QuickBooks or Stripe and drafts the reminders, prepares payroll in Gusto, and writes follow-ups from HubSpot deal history.
  • Pays through Stripe's Link wallet or Shop Pay using a virtual card, so a real card number is never handed over.
  • Keeps two actions off limits even with approval: blocks moving money between accounts and trading securities.
Three always-on assistants, three different distribution bets
AgentOperatorLaunchWho gets it nowReach
Claude DuetAnthropicOct. 21, 2026US Pro and Max subscribers, waitlist6,000+ Claude Marketplace connectors
DotsOpenAISept. 29, 2026ChatGPT Pro and Business Premium4,000+ apps inside ChatGPT
MuseMetaSept. 8, 2026US only, free with a usage limitDedicated secure VM; connector count not disclosed

Anthropic was candid about the edges: the agent can misread a thread, can misjudge what is urgent, is not yet strong at personal errands such as booking tickets, and cannot be watched live in its browser. Those limits are why this is a research preview rather than a launch, and why the financial detail investors want, waitlist size, cost per user and revenue per agent, has not been published anywhere yet.

The missing leg

The Bank Is One Permission Away, And It Is Not Anthropic's To Grant

Consumer bank accounts are the piece Duet does not yet have. Anthropic has been assembling it separately: a Money section appeared inside the Claude iOS app in mid-September with the prompt to link bank accounts and ask about spending, and by mid-October the internal name had shifted to Personal Finance, with Plaid integration visible in the app build. OpenAI has been there first. Since May 2026, ChatGPT's personal-finance preview has connected US subscribers to more than 12,000 financial institutions through Plaid, and Plaid's own account of that product describes granting read access with no write access.

Where Claude Duet is allowed to act, and which counterparty owns each rail
PermissionStatus in Claude DuetWho holds the rail
Read email and calendarAllowed, on a timerGmail, Google Calendar and other connectors
Read a business ledgerAllowedQuickBooks, HubSpot, Shopify and Stripe connectors
Pay a merchantAllowed with user approvalStripe Link wallet or Shop Pay virtual card
Read personal bank accountsNot in Duet; tested separately as Personal FinancePlaid
Move money between accountsBlockedNot applicable
Trade securitiesBlockedNot applicable

Before any big step, like sending a payment or deleting something, Claude Duet messages you for approval.

Anthropic, on Claude Duet, quoted by The Verge, Oct. 21, 2026
The read-and-pay-but-not-move design is the whole commercial argument. Duet can read a ledger and settle a purchase, yet it cannot shift a cent between accounts or touch a brokerage, so the bank earns nothing until a regulator says otherwise, while the aggregator and the card network collect on every connection and every transaction.

What is actually scarce

The Moat Is a Permission Graph, Not a Model

Model quality is not the moat in this product. Anthropic says Opus 5.5 performs at the level of Claude Fable 5.1 while costing 40% less to run than Opus 5, which is exactly the kind of price-performance convergence that erodes model-based switching costs quickly. What is scarce is permission: the right to read a ledger, to hold a card credential, to see a bank balance. Anthropic does not own those permissions, it rents them from Plaid, Stripe, Intuit and Google, and each rival can rent the same ones from the same counterparties.

Anthropic's annualised revenue run-rate

Reported run-rate checkpoints ahead of the IPO filing. Anthropic went from a $9 billion run-rate at the end of 2025 to $47 billion by late May 2026.

Unit: $B annualised run-rate

Dec 2024

One billion dollars

1

Dec 2025

Nine billion dollars

9

Feb 2026

Fourteen billion dollars

14

Mar 2026

Nineteen billion dollars

19

Apr 2026

Thirty billion dollars

30

May 2026

Forty-seven billion dollars

47

That is the uncomfortable part of the bank-rail story. Plaid sits behind ChatGPT's finance preview, behind Meta's Muse and behind Alphabet's Gemini through Google Cloud and Google Wallet, and it passed $500 million in annual recurring revenue in Q4 2025 with growth near 40%. Every frontier lab can rent the same rail, so no frontier lab owns the bank connection. The defensible asset sits one level up, in the connector surface and the consumer relationship around it, and in the merchant records that only an accounting or payments vendor can supply.

Why a consumer agent, now

Claude's Problem Is Reach, and Duet Is a Metering Machine

The strategic logic of a consumer agent is a distribution problem rather than a capability problem. ChatGPT has roughly three times Claude's US paid subscribers and a 14-fold lead in monthly active users, and only about 4.5% of US consumers pay for any of the three assistants at all. Claude's mix skews to the top of the market: 7.3% of its payers sit on the $100-a-month Max plan against 1.1% at ChatGPT and 1.3% at Google, and the top 1% of AI payers average $903 a month while driving 19.5% of observed consumer AI spend.

Share of payers on each lab's priciest individual plan

a16z consumer panel, August 2026. Claude's payer base is the most concentrated at the top of the price ladder.

Unit: % of paying consumers

Claude, $100 Max

Claude consumers on the Max plan

7.3%

Google, top tier

Google consumers on its highest tier

1.3%

ChatGPT, top tier

ChatGPT consumers on its highest tier

1.1%

  • Duet is included at no extra cost, but its background checks consume the same allowance as regular chat, so idle time shows up as usage.
  • That metering turns an always-on agent into an upsell engine for the $100 Max tier rather than a free feature.
  • Because the top 1% of payers already average $903 a month, the revenue plan leans on the cohort least likely to churn on price.
  • The research preview keeps cost contained: US only, Pro and Max only, waitlist only, and no published waitlist size.

What the market has already decided

Software an Agent Reads Has Been Discounted; Software an Agent Runs Has Not

Market capitalisation change from each issuer's own fiscal year-end to the Oct. 22, 2026 close, with enterprise value to trailing revenue and the latest reported quarterly revenue growth
CompanyRole in the agent stackMarket cap changeEV / revenueLatest quarter revenue growth
IntuitQuickBooks and TurboTax now live inside Claude-59%4.4x+13.7% (Q4 FY2026)
HubSpotDeal history an agent reads to write follow-ups-41%3.5x+19.8% (Q2 2026)
SalesforceAccounts and pipeline inside Claude, beta-1%5.3x+10.8% (Q2 FY2027)
DocuSignSigns the documents an agent prepares+45%4.4x+9.4% (Q2 FY2027)
ShopifyShop Pay wallet the agent pays from+7%16.9x+33.7% (Q2 2026)
VisaSettles every agent-initiated purchase+13%17.4x+14.4% (Q3 FY2026)

The market has already made a call on which side of the agent divide a software company sits, and it is not a call on earnings. Intuit, whose QuickBooks and TurboTax now live inside Claude, lost 59% of its market value while revenue grew 13.7% in the June quarter, and its enterprise value to revenue multiple fell from 11.9 times to 4.4 times. HubSpot, whose customer and deal history an agent now reads to write follow-ups, is down 41% with revenue up 19.8%. The vendors that get executed on rather than read, DocuSign, Shopify and Visa, are up 45%, 7% and 13% from their respective fiscal year-end marks.

Causation is not proven here. Company-specific news moves these multiples, and Salesforce, which sells its own agents and also sits inside Claude, is flat. But the direction is consistent enough to work as the operating assumption for the next two quarters: being read by an agent is being discounted, being paid by an agent is not. That distinction is the real investable content of the Duet launch.

Supply chain

Who Actually Gets Paid: Compute, Data Rail, Payment Rail, Records

  • Compute, upstream: Anthropic's projected roughly $80 billion of cloud infrastructure spend through 2029 lands on accelerator vendors first, and Nvidia revenue grew 106% year over year to $96.22 billion in the July quarter.
  • Cloud, upstream: Anthropic books the full amount its cloud partners bill as revenue and parks their share in sales and marketing, so agent volume shows up as gross revenue for Anthropic and as infrastructure revenue for Amazon, Microsoft and Alphabet at the same time.
  • Data rail, upstream: Plaid sits upstream of the assistant rather than of the bank and sells the same connectivity into every frontier lab, with more than 12,000 institutions behind the link.
  • Records, downstream: Intuit supplies the ledger, HubSpot the pipeline, DocuSign the envelope and Shopify the storefront the agent reads before it acts.
  • Distribution, downstream: Salesforce put accounts, opportunities and pipeline into Claude in a Sept. 15 beta, which turns CRM data into agent input and agents into a new channel.

For the banks themselves the news is close to neutral, and that is the part the bull case gets wrong. Anthropic's agent can read a bank balance and pay a merchant but cannot move money or trade securities, and OpenAI's finance preview is read-only too. Card networks are the ones selling agent credentials: Visa's fiscal third-quarter disclosure counted more than 150 AI-powered applications and named Agent Score, an Agent Directory and a Token Assurance Framework, while Mastercard is scaling Agent Pay for Machines with more than 30 industry partners. The toll on agent-initiated commerce is being priced at the network layer, not the bank layer, and banks are not the beneficiaries investors may have assumed.

Horizons and risks

What To Watch From Here, And What Would Break the Thesis

  • Days to quarters: the public S-1, expected as soon as Friday, Oct. 23. Read three lines, consumer subscription revenue broken out from API, how much of the gross-versus-net presentation survives, and the agent compute commitments.
  • Days to quarters: Duet's exit from research preview, meaning pricing, geography beyond the US and whether Pro and Max limits are metered separately.
  • Days to quarters: whether Plaid-connected bank accounts ship inside Duet or stay a separate Personal Finance surface in the main app.
  • One to three years: whether seats stop compounding. If Intuit, HubSpot and Salesforce seat growth stalls while agent usage compounds, seat-based pricing gets rewritten as usage-based pricing.
  • One to three years: whether merchant relationships migrate to agents. If buyers arrive with no storefront, Shopify keeps settlement and loses the customer, and Visa keeps the interchange.
  • One to three years: whether Anthropic monetises the agent directly. On 2028 projections of $70 billion revenue and gross margin rising from roughly 50% toward 77%, agent usage has to become a priced line, not free background work.
Three things would falsify this read. If the public filing shows the same quarter as $11.5 billion in press coverage and $10.9 billion in Anthropic's own numbers, the gross-versus-net presentation is the real risk to the multiple, not the launch. If Duet converts waitlist signups into usage that does not lift Max mix, it is a cost centre. And if a rival ships equivalent bank connectivity first, the permission graph is a menu rather than a moat.

Still not disclosed

Claude Duet waitlist size

Not disclosed

No signup, activation or usage figures published

Personal Finance bank coverage

Not disclosed

Bank list, connection method and geography all unstated

Cost per Duet agent

Not disclosed

No unit economics on inference consumed per user

Anthropic consumer versus API revenue

Not disclosed

Consumer subscriptions sit inside the roughly 20% non-API share

Agent-commerce revenue share

Not disclosed

No economics published for Shop Pay or Stripe Link agent payments

Bank-side economics of agent access

Not disclosed

Plaid discloses annual recurring revenue, not per-agent pricing

The investable chain behind Claude Duet

IIntuitINTU--
--Vol --
-
Mixed
  • QuickBooks and TurboTax already sit inside Claude, so the ledger becomes agent fuel without a seat
  • Q4 FY2026 revenue grew 13.7% to $4.35B, so the 59% market value loss is a model call, not an earnings call
  • Horizon: if agents read ledgers faster than humans add seats, QuickBooks pricing moves to usage within one to three years
HHubSpotHUBS--
--Vol --
-
Bearish
  • Deal history is a named Duet input, and the pipeline becomes readable without a logged-in user
  • EV/revenue has already fallen from 6.5x at the Dec. 31, 2025 year-end to 3.5x, with Q2 2026 revenue still up 19.8%
  • Horizon: no transaction rail defends the seat, so seat-based CRM pricing is exposed over one to three years
DDocuSignDOCU--
--Vol --
-
Bullish
  • Duet signs documents once a user approves, and agent-initiated signatures are incremental volume
  • Market capitalisation is up 45% from the Jan. 31, 2026 fiscal year-end with Q2 FY2027 revenue up 9.4% to $876M
  • Horizon: envelopes drafted, approved and signed by an agent convert into volume over one to three years
SShopifySHOP--
--Vol --
-
Mixed
  • Shop Pay is one of two wallets Duet pays from, and Q2 2026 revenue grew 33.7% to $3.58B
  • Market value is up 7% from the Dec. 31, 2025 year-end while read-side software peers fell 41% to 59%
  • Horizon: keeps settlement but loses the customer relationship if the agent becomes the storefront's buyer
CSalesforceCRM--
--Vol --
-
Mixed
  • Accounts, opportunities and pipeline entered Claude in a Sept. 15 beta, so it is hedged between selling agents and feeding them
  • Q2 FY2027 revenue grew 10.8% to $11.35B and EV/revenue is roughly flat at 5.3x
  • Horizon: distribution through a rival model protects seat pricing only if agents also carry Salesforce's own monetisation
VVisaV--
--Vol --
-
Bullish
  • Every agent purchase settles on its rails, so Visa collects on agent-initiated checkout
  • Fiscal Q3 2026 net revenue grew 14.4% to $11.63B while the release counted more than 150 AI-powered applications
  • Horizon: agent credential products such as Agent Score and an Agent Directory are the one to three years toll mechanism
NNvidiaNVDA--
--Vol --
-
Bullish
  • Background checks are inference work, so an always-on agent turns idle hours into token demand
  • Revenue grew 106% year over year to $96.22 billion in the quarter ended Jul. 26, 2026
  • Horizon: Anthropic's projected roughly $80 billion of cloud spend through 2029 is the near-term order read-through
GAlphabetGOOGL--
--Vol --
-
Mixed
  • Gmail is a named Duet connector and Alphabet is an Anthropic cloud supplier, so it supplies the inbox the agent reads and competes for the same slot
  • Q2 2026 revenue was $119.80 billion, with Gemini reaching bank data through Google Cloud and Google Wallet
  • Horizon: if agents own the inbox, the productivity suite's seat becomes the asset at risk over one to three years