Anthropic shipped something on November 17 that answers a question every agency and SaaS builder on the Claude API eventually runs into: who actually pays when a customer uses the AI tool you built them? The honest answer is "it depends how your customer buys Claude" — and for most small businesses right now, the answer is still you.
What actually launched
The feature is called linked billing, live in beta under the header user-profiles-billing-2026-10-01. It lets a customer pay for the Claude usage inside your app through their own Claude organization, instead of you marking up every token after the fact. A customer links their org to your app once. From then on, you set billing_policy on each request: bill_to: "self" keeps the charge on your account, bill_to: "user_profile" sends it to theirs. Leave the field off and nothing changes — it bills you, same as every request has worked until now.
It's the first time Anthropic has let a third-party app point an individual API call at someone else's wallet.
The line that cuts out most small businesses
Here's the part that matters more than the feature itself. A customer only qualifies if their Claude organization pays Anthropic directly — prepaid Console credit or an invoiced account. Anyone on a Claude subscription plan, Pro, Max, or Team, gets refused outright, with a 403 and the error code user_profile_not_billable. So does anyone billed through the AWS, Google Cloud, or Azure marketplace instead of Anthropic's own billing. That line cuts out most of the small businesses buying AI through somebody else's app today. If your client runs Claude Max for their team and nothing else, linked billing has nothing to offer them. Not yet.
Who this is actually for
Realistically: agencies and SaaS vendors selling into clients who already have a direct Anthropic billing relationship for something else — an invoiced account, an internal AI platform team, a company large enough to negotiate its own API terms. If you build custom AI agents for clients and one of them already pays Anthropic directly for a different project, this is worth a real conversation. If you're a five-person service business paying $20 a month for Claude Pro, you're not the audience. You might be eventually — this is beta, and Anthropic says the eligibility rules can still move.
The compliance catch nobody mentions
Even where linked billing works, the request still runs on the app builder's organization, not the customer's. The builder's rate limits, model access, retention settings, and encryption keys apply. Not the customer's. A client who has zero data retention or a signed BAA for their other Claude usage doesn't carry that protection onto a linked-billing request running through your app — the documentation is explicit that those arrangements don't transfer. If you're an agency with a healthcare or legal client whose compliance terms depend on that coverage, don't route that traffic through someone else's linked billing connection.
The part actually worth using
There's a genuinely useful angle buried in the mechanics, though. Linked billing only moves the raw token cost. It doesn't stop you from invoicing a separate fee for the build and the upkeep. That's a cleaner split than most agencies offer right now — the usual approach is estimating a client's monthly usage, adding a margin, and billing one blended number the client has no way to check. Once a client's org is eligible and linked, you bill your own fee for the work, and Anthropic's own usage reporting shows them exactly what their AI cost that month: your dashboard shows a zero charge, theirs shows the real number. Clients trust a figure they can see more than a figure you tell them.
What to actually do with this
For nearly everyone reading this, nothing changes this week. Subscription-billed organizations can't link yet, this is still beta, and setting it up on the builder's side, registering an OAuth app, standing up a service account, signing your own tokens instead of using an Anthropic-issued secret, is real engineering work, not a settings toggle. The one thing worth doing now: if you bill clients for an AI feature by estimating their usage and tacking on a flat markup, start noting which of those clients already pay Anthropic directly for something else. Those are the two or three who'll qualify the day this leaves beta, and "you pay Anthropic for what you use, you pay us for building and running it" is a better pitch than a blended number nobody can audit.
We've written before about what an AI agent actually costs to run month to month, and the retainer-versus-one-off question that usually comes up right after. If you're trying to figure out how to bill AI usage into a client project, or whether building the agent is worth it in the first place, that's exactly what we scope on a discovery call.
— Cole
Sources
- Claude Platform release notes — linked billing beta announcement, verified November 18, 2026
- Organization billing connections documentation — eligibility rules, billing_policy mechanics, and error codes