Field noteNov 20, 20266 min read1 source

Anthropic’s Claude Platform documentation describes Linked billing

In one minute

Anthropic’s Claude Platform documentation describes Linked billing, an OAuth flow that lets a product’s users authorize Claude API usage to be billed to their own Claude organization. For teams embedding Claude in a product, it separates customer-directed usage from the developer’s own bill and specifies what happens when customer billing cannot be applied.

What happened

Anthropic’s Claude Platform OAuth documentation says its Linked billing feature lets a product charge Claude API usage to each customer’s own Claude organization. The intended use is a product that calls Claude on a customer’s behalf but does not want every such call to be paid from the product developer’s API account. The customer starts by clicking a “Link your Claude organization” control inside the product. They leave the product for Claude’s approval page, sign in to Claude, approve the app, and are returned to the product with a one-time code.

The customer needs either the Admin or Billing role in their Claude organization. Anthropic says the approval page displays the app’s name and the permissions it is requesting, giving the user a chance to see what connection they are authorizing. The product back end receives the one-time code and exchanges it for a connection ID. That connection is the basis for later customer-billed requests. Customers can unlink the application at any time through the controls Anthropic provides.

After a connection exists, the product developer chooses for each Claude API request whether the customer pays or the developer pays. Anthropic specifies a consequential failure mode: when a request is marked for customer billing but cannot be billed to that customer, the Claude API rejects the request rather than falling back to the developer’s account. The documentation also sets a usage boundary. A product may bill a customer only for requests the customer directs it to make; it may not use the connection to charge the customer for the developer’s own usage.

Anthropic assigns different responsibilities across the flow. It operates the approval page and the Claude Console page where a customer can set a spend cap or unlink an app. The product developer remains responsible for support questions about the product itself. This means the integration is not simply a payment flag in an API call: the developer must build an understandable linking experience, decide which user actions create customer-billed requests, and support the product behavior around those choices.

The documentation describes one-time configuration in Claude Console. An administrator creates an OAuth app, copies its client ID, and configures the app name and return address presented during approval. The back end does not use a conventional client secret. Instead, it proves it is the app using a short-lived token from the developer’s identity provider. Anthropic names AWS, Google Cloud, and Kubernetes as examples of possible identity providers, and the developer selects which provider tokens the app should accept.

The setup also requires a service account and a second federation rule through Workload Identity Federation. The developer selects that service account as the OAuth app’s operator service account, which the back end uses in the customer-billing path. Linked billing is limited to organizations that pay Anthropic directly for API usage. It does not work for organizations billed through a cloud marketplace. Anthropic points developers to separate Linked billing setup guidance and OAuth reference material for endpoints, parameters, token responses, token lifetimes, and errors. The overview does not state a release date, a price, or a version number.

What the archive adds

We have not covered this before; it opens a new thread.

Trend signal

This does not yet fit a pattern New Runtime tracks. The feature creates a specific authorization and billing arrangement between an application, its customer, and Anthropic, but this documentation alone does not demonstrate a broader change in agent interoperability, capability exposure, or runtime security boundaries.

Why it matters

For teams building AI products, who pays is part of the product boundary. A product may need to offer Claude-powered work to organizations that already hold their own Anthropic relationship, budgets, and internal approval process. Linked billing gives such a customer a way to authorize use of its own organization without handing the product developer an API key. It also lets the developer retain control over which individual API requests are customer-billed and which remain part of the developer’s own operating cost.

The explicit rejection behavior is especially important. A system that silently charged the developer whenever customer billing failed could turn a configuration error, an ineligible organization, or a disconnected account into unplanned variable cost. Anthropic’s documented behavior instead makes failed customer billing visible as a rejected request. Teams will still need product decisions for that event: whether to ask the user to reconnect, offer a developer-paid fallback where appropriate, or stop the task and explain why.

The feature has meaningful operational prerequisites. A customer needs the required organizational role and direct Anthropic API billing. The developer needs OAuth app configuration, an identity-provider token arrangement, a service account, and workload identity federation. That makes Linked billing more suitable for products with a real backend and clear customer-account model than for a quick client-side integration.

Limits and uncertainty

Anthropic’s overview explains the authorization flow, eligibility rules, and required setup, but it does not report adoption, customer outcomes, or measured reductions in billing complexity. It does not state a launch date, pricing, request-level fees, or any cost of OAuth app administration. The page says linked billing is unavailable to organizations billed through cloud marketplaces, but it does not quantify how many prospective customers that excludes. It also describes the technical controls without independently measuring how teams handle billing disputes, mistaken user actions, interrupted authorization, or support incidents in production. A successful connection is not itself proof that a product’s payer logic or customer communication is appropriate.

Skill to build now

Map one Claude-powered user action in your product to its intended payer before implementing the connection flow. Write down when the request is customer-directed, when the developer should pay, and what the interface does after a customer-billed request is rejected. Then test the OAuth return path, unlinking path, and spend-cap explanation with an administrator or billing-role user. This turns the per-request payer choice into an explicit product rule rather than an undocumented cost assumption.

Reader takeaway

Linked billing is a concrete option for a Claude-powered product whose customers pay Anthropic directly and want their own organization charged for customer-directed API usage. The implementation question is not only OAuth setup: design the payer decision per request, plan for hard rejection when customer billing fails, and make connection, spend-cap, and unlinking controls intelligible to the people authorized to use them.

Evidence / sources

  1. 1platform.claude.com/docs/en/manage-claude/oauth-apps/overview
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicClaudeExplore the claude topic hub.
  2. 02topicAnthropicExplore the anthropic topic hub.
  3. 03archiveField NotesOpen the latest editorial analysis.
  4. 04sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract