Article

Claude linked billing: who can pay for your app?

For app teams, check which Claude accounts and credits qualify, how spending caps work, and why customer billing does not move data controls.

Editorial illustration for Claude linked billing: who can pay for your app?: a geometric block represents a model release. Not documentary evidence.

Anthropic’s November 17, 2026 linked-billing beta lets software businesses charge eligible Claude API usage to a customer’s own Anthropic account. It works for directly billed API organizations—not subscription-only or cloud-marketplace accounts—and the app operator’s data settings still apply, under Anthropic’s eligibility rules.

Customers authorize a connection instead of handing the app their API key. The launch announcement describes a choice of payer for each request, not free Claude access or a transfer of the customer’s whole Claude environment.

Which customers and credits qualify?

Linked billing needs eligible API funding, not just a Claude subscription. The customer’s API organization must pay Anthropic directly: purchased Console credits and invoiced billing qualify. Organizations billed through AWS, Google Cloud or Azure marketplaces do not. These are account-level eligibility rules; an employee’s paid Claude plan alone does not establish eligibility.

Monthly API credits included with Max, Ultra and Team plans cannot pay for linked third-party usage. Anthropic’s linked-billing documentation also currently excludes partner, venture-capital and startup-program credits. Do not budget from the displayed credit balance without checking which credits can fund the app; the help center says credit-program rules may change.

A subscriber can still use a separately eligible, funded API organization. Claiming Max and Team subscription API credits is a different connection from authorizing a third-party app to bill that organization. The included credits do not become eligible simply because both connections exist.

Which features can the customer pay for?

The beta supports customer billing for the Messages and Chat Completions APIs. The app chooses who pays for each supported request. Three exclusions matter when deciding whether the whole product can use this funding model:

  • Batch processing: Message Batches cannot charge the linked customer. A customer-paid batch entry is rejected without charging either party; other entries continue. Run batch work as operator-paid if you need it.

  • Billable code execution: a customer-paid Messages request using these features is rejected. It must be sent as operator-paid to run. This restriction is specifically about billable execution, not every use of code execution.

  • Managed Agents: this service is not yet covered by linked billing.

These limits come from the endpoint documentation and linked-billing help page. A product may therefore need mixed funding: customer-paid interactive replies alongside operator-paid batch work. Explain that split before asking customers to connect.

Only customer-directed work should be charged to that customer; the operator’s own evaluations and work for others are excluded by Anthropic’s attribution rules. Budget separately for those activities, app hosting and any software subscription fee. Changing who pays for model usage does not establish a saving.

Which data settings apply?

The requests still run in the app operator’s organization. Its model access, rate limits, service tier, retention settings and encryption keys apply. Linked customers share the operator’s capacity limits; their own settings do not travel with the bill, as the launch article and detailed guide explain.

This is a separate purchasing decision from funding. A customer that can pay is not necessarily a customer whose data requirements fit. Anthropic excludes customers needing US health- or education-record coverage under HIPAA or FERPA, or customer-managed encryption keys, during the beta. Their own zero-retention arrangement or healthcare business associate agreement does not govern these requests. These are Anthropic’s stated beta restrictions, so check the intended data use before offering the connection.

How do monthly spending caps work?

Customers manage app caps in Console under Linked platforms. An app cap and the organization-wide monthly spend limit are separate: reaching either stops customer-billed requests. A $0 app cap blocks all such requests. Anthropic’s cap rules also say a default cap applies independently to every app without its own override.

Illustrative budget: a $100 default across three apps permits up to $300 in combined monthly cap-counted usage—not $100 shared between them—assuming sufficient eligible funding and no lower organization-wide limit. An app-specific cap replaces the default for that app.

Caps count token costs and web-search fees at list prices before negotiated discounts. They are not necessarily the amount on the customer’s invoice. Cap-based refusals continue until the next month starts at 00:00 UTC, or the customer raises or removes the applicable cap. Anthropic says changes can take about a minute; they are not instantaneous. These details are in its spending-cap documentation.

If a customer-paid request cannot be billed, Anthropic refuses it instead of silently charging the operator. An operator can deliberately resend it at its own expense, but that is a separate spending choice. A billing error can reflect an exhausted eligible balance or a limit, while a permission error does not uniquely mean the customer unlinked. Build an actionable account-and-billing message around the documented failure cases, not an automatic subsidy or endless retry.

What must be ready before customers connect?

An Admin or Billing member approves on the customer’s behalf. The link belongs to the organization and remains active if that employee leaves. New apps start in testing, with up to 10 listed test users and mandatory monthly caps of $50 or less. The setup guidance makes testing distinct from a public rollout.

Operators with invoiced billing can make an app live when they confirm publication. Others must verify domain control and pass Anthropic review. Existing test caps remain after the app becomes live until customers change them. Follow the app-publication instructions rather than promising immediate access to every customer.

Before rollout, settle which customer organization each app account bills. Within one app, one customer organization links to one billing user profile. The standard operator credential can charge any linked customer, so a wrong profile mapping can charge the wrong organization without a billing error. Anthropic documents this customer-attribution responsibility; authorization alone does not validate your customer mapping.

Keep request-level records of who was meant to pay. Operator reports include usage from both payers but do not reveal the customer’s actual charges or remaining app cap. Reporting is delayed, so it is not a live budget control. RohitAI’s guide to Claude usage and cost analytics explains the related reporting tradeoffs; the linked-billing monitoring guide defines what the operator can see here.

Also plan for revocation. Customers can unlink in Console, but the app receives no unlink notification. Anthropic says new calls usually stop within about 10 seconds; calls already running finish and are billed. A connection marked linked does not itself prove that requests can be funded. These are documented connection limits, not guarantees of instant enforcement.

When is linked billing a good fit?

Based on these rules, it is worth evaluating when customers already have directly billed API organizations, eligible funds, supported workloads and data requirements that fit the operator’s arrangements. A subscription-only audience, dependence on excluded credits or unsupported features calls for separate funding or a mixed design. A requirement for the customer’s own data protections cannot be solved merely by moving the bill.

Start by qualifying one intended customer’s funding and data requirements, then map every product feature to its payer before implementing the connection. Keep the beta status in the decision: Anthropic warns that its behavior can change and may not suit production use.