Claude API Adds Linked Billing So Customers Pay Anthropic, Plus Self-Hosted Managed Agents

Linked billing (beta) charges a customer's own Claude org per request, but the request runs in yours. Self-hosted Managed Agents is a private beta.

Anthropic has made two changes in the past week to who pays for Claude and who hosts it. Linked billing, launched in beta on November 17, lets a customer connect their own Claude organization to your app once, after which any request whose billing_policy you set to bill_to: "user_profile" is charged to that organization instead of yours. Self-hosted Claude Managed Agents, released on November 12, puts the Managed Agents service and its sandboxes on your infrastructure, where you store agents, sessions and session history. The release notes say to request access through your Anthropic account team, and Anthropic’s launch post calls it a private beta. Release notes · Linked billing post · Self-hosted post

Everything below is vendor-reported. It comes from Anthropic’s documentation, launch posts and self-hosted license, read on November 17. S5 Labs has not built against linked billing or run the self-hosted server. The only outside coverage we found, from Help Net Security and AI Weekly, restates Anthropic’s self-hosting post.

How linked billing works

Setup goes through an OAuth app, which a Console member with the Admin role creates. For each customer you create a user profile, send someone from their organization to an approval page on auth.claude.com, and exchange the returned code, which is single-use and expires after 5 minutes, for a connection ID. The docs say the approver needs the Admin or Billing role; Anthropic’s post also lists Owner. Linked billing docs · Linked billing post

After that, you choose billing per request. Anthropic’s example, sent with API version 2026-10-21, adds this to the body:

"billing_policy": { "type": "billing_policy", "bill_to": "user_profile" }

The customer’s profile goes in the anthropic-user-profile-id header, next to the beta header user-profiles-billing-2026-10-01. With bill_to: "self" or no billing_policy, you pay as before. Customer billing works on /v1/messages and /v1/chat/completions, and /v1/messages/count_tokens accepts the field and ignores it. Eligible customers and endpoints

Neither the docs nor the post mentions a fee or markup. The post says customers “can then draw down from their existing Anthropic commitment, or pay as they go”. You see each customer’s request count, token usage and models in the Console or the Usage and Cost Analytics API, but the docs say you cannot see what the customer paid. On those requests you stop reselling tokens, so you carry no inference cost and earn no token margin. Linked billing post · Linked billing docs

Who can be billed

The customer’s organization has to pay Anthropic directly for its API usage.

Customer’s situationResult
Pays Anthropic directly, with prepaid credits or invoiced billingBillable
Billed through the AWS, Google Cloud or Azure marketplace403 user_profile_not_billable
Has only a Claude subscription plan403 user_profile_not_billable
Has only plan credits (the monthly API credits with Max, Ultra and Team) or partner, VC or startup program creditsThose credits can’t pay, and the API can return a 402

Eligible customers and endpoints · Which credits can pay

A Claude subscription alone doesn’t make a customer billable, and plan credits can’t pay. (OpenAI began bundling monthly API credits with ChatGPT Pro and Business plans in October.) You also can’t check eligibility from your side; the docs tell you to ask your account team. A completed link doesn’t prove billability either, so they say to word your success message as “linked” rather than “billing active”. The billing page covers only the first-party Claude API and does not say whether linked billing works on Claude Platform on AWS or other cloud platforms. Linked billing docs

The sharp edges

The request still runs in your organization

Your rate limits, model access, service tier, data retention and encryption keys apply to a customer-billed request. The customer’s zero data retention (ZDR), BAA and customer-managed encryption key (CMEK) arrangements do not, and the launch post adds IP allowlists. The docs say it outright: “Don’t link customers that need HIPAA, FERPA, or CMEK coverage for this usage.” Linked billing docs · Linked billing post

An unbillable request is refused, not charged to you

The docs: “Anthropic either bills your customer or refuses the request. It never silently bills you instead.” These are the refusals to plan for:

ResponseWhat it meansRetry
403 user_profile_not_billableThe customer can’t be billed, and the response doesn’t say whyOnly in the first 10 seconds after a new connection
402 linked_organization_spend_limit_reachedA spend cap or the organization’s monthly limit was reached, or no credit can pay. Refused before the reply startsNo
402 linked_organization_billing_unavailableAnthropic can’t currently bill the organization, for example because it is suspended or its credit ran out after Claude started a reply. The docs don’t say who pays for a partial replyNo
503 overloaded_errorBillability couldn’t be confirmed. Nothing was served or billedYes, with backoff, honoring retry-after

On /v1/messages the code sits in error.details.error_code. /v1/chat/completions returns the same status and message in OpenAI format without the code. Any of the first three can be served at your expense by resending with bill_to: "self" from your back end, using the operator service account’s own credential. Make that a product decision for each error code. A blanket retry turns a customer’s $0 cap into your bill. Expected errors

The wrong header bills the wrong customer

By default, the operator service account’s credential can bill every customer linked to the app. The docs: “If a request with that credential names a different profile linked through this app, Anthropic bills that profile’s organization, and no error tells you.” A successful response doesn’t say who was billed, so the docs tell you to keep your own record. Your customer pays · Expected errors

The safer mode is a per-connection access token, which can bill only that connection’s customer and expires within an hour, combined with a setting that turns off operator-only authentication. Neither is self-serve in the beta: token requests fail until they are enabled for your organization, and the setting can’t be turned on until it is available to your organization. The docs say to ask your account team about both. Tokens add a failure mode of their own: “A used refresh token that arrives again more than 30 seconds after its first use ends the connection”, and the customer has to link again. OAuth tokens for each connection

What stays on your bill, and what you can’t see

Message Batches does not support customer billing. A batch item marked bill_to: "user_profile" errors, nobody is billed for it and the rest of the batch runs. A customer-billed request that uses the billable code execution features returns a 400, a known limitation per the docs, so send those as bill_to: "self". Eligible customers and endpoints

The customer can set a monthly cap for your app in whole US dollars, as low as $0, and you can read neither the cap nor the usage against it. The cap counts token costs and web search fees at list prices, before any negotiated discount. Once it is reached, refusals last until 00:00 UTC on the first of the next month unless the customer raises or removes it. Customers can unlink at any time, and the docs warn that “you get no webhook event and no email.” Until your app is live, every link needs a cap of $50 or less. Spend caps · After linking

What self-hosting Managed Agents involves

Anthropic’s docs now describe three ways to run Managed Agents:

OptionSandboxesServiceModel
Cloud sandboxes (default)AnthropicAnthropicAnthropic
Self-hosted sandboxesYouAnthropicAnthropic
Self-hosted Managed AgentsYouYouAnthropic, or a provider that you choose

Self-hosted overview

The server, cma-server, serves the API, runs the agent loop and stores history. Its database is PostgreSQL 16 or later, and without one the data lives in memory and a restart erases it. Workers poll the server and run the shell and file tools in your sandboxes, and a web console is served on the same address and port as the API. The Docker Compose stack “is for development, not for production”, and for production Anthropic recommends the release’s two Helm charts. Model calls go to the Claude API by default, or to Amazon Bedrock, Claude Models on Google Cloud (Google managed), Microsoft Foundry or your own gateway. Self-hosted overview · Quickstart · Inference provider · Data and access

The server has the same API as the hosted service, with a long list of exceptions. Cloud environments, the Files API, webhooks, scheduled deployments, dreams and mcp_oauth vault credentials are all missing, and the SDK’s /v1/messages and /v1/models calls return 404. Nothing is copied over from the hosted service, session history included. The server records no traces or metrics until you enable OpenTelemetry, which we covered for the hosted service in October. Differences from the hosted service · Data and access

What still leaves your network

Self-hosting moves storage to you. Inference still happens at the provider, and per the Data and access page, “Each model call carries the conversation so far, including prompts, tool inputs, and tool results.” When the provider is the Claude API, calls also carry up to three headers holding the session, agent and environment IDs. The docs say the IDs hold no name or content but encode a creation time, and “No setting turns them off.” Requests to the Claude API also name the server software and version in the User-Agent, which the docs say Anthropic can use to count usage of self-hosted Managed Agents. Going the other way, “Anthropic has no remote access to the server or its log”, and the server connects to Anthropic only when the Claude API is its inference provider. Data and access

The docs say a ZDR arrangement covers these model calls the same way as other Messages API calls, and the transcripts sit in your database. Hosted Managed Agents is not ZDR-eligible, because Anthropic stores the transcripts. The self-hosted exception is for ZDR only: the data-retention page still lists Managed Agents as “Not eligible for HIPAA”, sub-features included. Deleting a session removes it from your database but not from backups, exported logs and traces, or memory stores. Data and access · API and data retention

Isolation is your job

The self-hosted overview warns: “The agent’s bash tool is a real shell with the worker process’s permissions and network access. The worker doesn’t isolate that shell.” The Compose stack and the cma-sandboxes Helm chart start each session’s worker in its own container or pod. (Claude Code’s sandbox is a separate product; see our Claude Code article.) Self-hosted overview

Tenancy is coarse. “Environments in one deployment don’t keep trust levels apart”, so the docs prescribe one tenant and one trust level per deployment. The server API key reaches every environment, and code the agent runs can call the server with any credential in its sandbox, so keep the environment key out of it. Data and access

Access, price and license

We found no self-serve path and no general availability date. The pricing page says only: “For pricing of self-hosted Claude Managed Agents, contact your Anthropic account team.” The license grants a royalty-free right to run the binary and leaves model calls under your existing Anthropic agreement, which is different from saying the product is free. It binds you from the first download and sets deadlines someone has to staff:

ObligationDeadline
Deploy a Safety Update1 business day after delivery
Disable a feature or roll back a version on Anthropic’s written notice8 hours
Report an Agent Incident48 hours after confirming it
Terminate for convenience, either side90 days’ notice

The license also bars modifying the code or giving it to third parties, customers included (affiliates and contractors acting on your behalf are excepted), and caps Anthropic’s total liability at US$100,000. You may not remove, block, alter or falsify any usage signals the software sends. Anthropic may decide, for safety reasons, that some models are reachable only through its hosted service. The license doesn’t say which, so ask before planning around a specific model such as Fable 5.2. License · Pricing

Which product fits, and what to check

Linked billing fits a SaaS product whose customers already hold direct-billed Claude Platform organizations and accept that their traffic runs under your data terms. It is the wrong tool for customers who need HIPAA, FERPA or CMEK coverage or their own ZDR, for marketplace-billed customers and for batch-heavy workloads. Anthropic says the beta “might not be suitable for production use”. With the token mode gated, your own code is the only check against billing the wrong customer, so we would ask for token access before building on it.

Self-hosted Managed Agents fits a team that must keep agents and session history on its own infrastructure, or must send inference through Bedrock, Google Cloud, Foundry or a gateway, and that can run Postgres, Kubernetes and sandbox isolation itself. If the requirement is only that tools run on your hardware, the docs point to self-hosted sandboxes, where Anthropic still runs the service.

For a security review:

  • Profile mapping. Derive the profile ID on your server from the authenticated tenant, and log every request sent with bill_to: "user_profile" together with its request-id header.
  • Disclosure. A customer paying from their own organization could assume their Anthropic terms come along. Say at linking that your retention and encryption settings govern their requests.
  • Tenancy. Plan one self-hosted deployment per tenant and trust level, and keep environment keys out of sandboxes.
  • Egress. Write down what leaves: full conversations to the inference provider, up to three ID headers on the Claude API path, and usage signals you may not block.
  • Contract. Get the self-hosting price in writing, and have counsel read the 1-business-day and 8-hour deadlines against your change-control process.

Sources

Anthropic announcements and license

Developer documentation

Secondary coverage

Continue reading.

Insight7 min read

Claude Managed Agents Adds Server-Run Workflows and OpenTelemetry Tracing

Anthropic's Managed Agents can now run fan-out workflows on its servers and export traces over OTLP. Limits, billing, content defaults and what to test.

Insight3 min read

Claude Opus 5.5 Cuts Costs and Changes Conversation Handling

Anthropic's Opus 5.5 release lowers token prices. Its preserved-thinking rules also make conversation history a migration concern for custom agents.

Insight5 min read

Claude Sonnet 5: Anthropic Cuts the Price and Closes the Gap to Opus

Claude Sonnet 5 ships at $2/$10 launch pricing as the most agentic Sonnet yet. The real story: the gap to Opus 4.8 and a new tokenizer.