Back to Explore
NewsVibeFix editorialUpdated Nov 21, 2026

Anthropic Launches Claude Linked Billing: Use Claude Inside Third-Party Tools Without Handing Over Your API Key

On 17 November 2026 Anthropic launched linked billing in beta on the Claude Platform: link your Claude organization to a third-party platform once, then choose per request whether the platform or you pay — no API key changes hands. But requests still run in the platform's organization, so your data terms do not travel with them.

Anthropic's official illustration: the Claude Console 'Linked platforms' page, showing each platform's monthly cost and spend cap

In short: from now on, using Claude inside a third-party tool can be billed straight to your own Anthropic account

On 17 November 2026, Anthropic launched linked billing in beta on the Claude Platform. The idea is simple: a customer links their Claude Platform organization to a third-party platform once, and from then on the platform can choose, on each Claude API request, whether the platform pays or the customer's organization pays directly. Customers no longer have to hand their API key to a third party, and platforms no longer have to front the model cost for their customers. Anthropic's announcement names the target scenarios clearly — coding agents and workflow automation tools, the kind of product where vibe coding users make heavy daily use of Claude. Nineteen platforms already support it at launch, including Factory, Linear, Ramp, Warp, Railway, Kilo Code, and Harness.

To see why this matters, look at the problem it solves. Until now, using Claude inside a third-party tool usually meant one of two ways to pay: either the platform folded the model cost into its subscription price and paid for you, or it asked you to paste your own Anthropic API key into it. The first way makes pricing hard to get right for the platform; the second way is worse for the user: an API key works like a password, and anyone who has it can spend your money in your name. The key has to live on the platform's servers, be encrypted, be rotated, and if it leaks, the loss is direct and financial. Many small teams would not dare ask users to paste a key, and many users would not dare paste one — this feature exists to take that wall down.

The linked billing flow has three steps. First, the customer clicks "Link your Claude organization" inside the platform, jumps to Anthropic's own approval page at auth.claude.com, signs in, picks an organization, reads how billing works, sets a monthly spend limit, and clicks "Link and authorize". Second, the platform adds a billing_policy field to each request, saying whether this one is paid by the platform (bill_to: "self") or by the customer (bill_to: "user_profile"), together with the user profile's ID. Third, Anthropic bills the customer's organization directly, and the customer can see in Console how much each linked platform spent, on which models — and can unlink any platform at any time. At no point in the whole process does an API key change hands.

Anthropic also gives platforms two useful capabilities. One is per-customer usage visibility: a platform can look up each linked customer's request count, token usage, and models used in Console, or programmatically through the beta Usage and Cost Analytics API. The other is per-request payer selection: for the same customer, a platform can cover some features itself and pass others to the customer — basic features on the house, advanced features on the customer's bill. That "choose who pays on each request" design is something the API-key model could never do: with a key, one key maps to one billing owner, and the payer is fixed. Linked billing turns the payer into a parameter you can set on every request.

Flow diagram: the customer links their Claude organization and sets a monthly cap, the platform chooses the payer on each request, and Anthropic bills the customer's organization directly; requests still run inside the platform's organization

But there is a big precondition: requests still run inside the other side's organization

The easiest place to misunderstand this feature is also the place Anthropic stresses most in its announcement and documentation: linked billing changes who pays, not where requests run. Requests still execute inside the third-party platform's own organization, under the platform's rate limits, model access, service tier, data retention, and encryption keys. The customer's own organization settings — IP allowlists, customer-managed encryption keys, zero data retention, business associate agreements — do not travel with the request. The documentation says it plainly: do not link customers that need HIPAA, FERPA, or customer-managed encryption key coverage for this usage.

In plain language: linking the bill does not carry your compliance protections with it. If your company has hard requirements on data retention, or your industry requires a business associate agreement, then using linked billing inside a third-party tool means your data is still handled under the platform's rules, not your company's rules. That is not an oversight by Anthropic; it is a built-in boundary of the design: the request runs in the platform's organization, so the platform's rules decide. Before adopting it, read the platform's own data processing terms first, then decide which work can go through linked billing and which must stay as direct calls inside your own Claude organization.

There is a related limit: the platform can see each linked customer's usage. Anthropic says platforms can see each customer's request count, token usage, and models in Console. That is good for the platform's bookkeeping, but for the customer it means your usage details are visible to the platform. Before linking, think about whether you mind the platform knowing how many Claude tokens you burn there each month. For most everyday development use that is nothing; but if usage privacy is something you care about, settle that question first.

Not everyone can use it: eligibility, credits, and how errors work all have fine print

Linked billing is not open to every Claude user. Only customers whose Claude Platform organization pays Anthropic directly can be billed this way: prepaid credits and invoiced billing both count. Organizations billed through the AWS, Google Cloud, or Azure marketplaces cannot use it, and neither can organizations with only a Claude subscription plan — requests for them are refused outright with a 403 user_profile_not_billable. Which credits can pay also has rules: the monthly API credits that come with Claude Max, Ultra, and Team plans cannot pay for these requests, and neither can credits from Anthropic's partner, venture capital, and startup programs. Only credits bought in the Claude Console, or invoiced billing, can pay.

The error behavior deserves to be called out on its own. Anthropic's documentation states that if a request marked for the customer cannot be billed to them, the Claude API refuses it rather than billing the platform. When there is no money or the spend limit is hit, the response is a 402 — and the response does not say whether the cause was a lack of credit or a spend limit. That is deliberate, to avoid leaking the customer's account state. For platform developers, the direct benefit is that you will never silently front a customer's bill without knowing. But there is a cost to handle: your code must deal with that refusal — for example, showing the user a clear message and guiding them to raise the limit or top up — instead of letting the request fail quietly in the background.

The technical coverage has a few gaps too. billing_policy is currently supported on three endpoints: /v1/messages, /v1/chat/completions, and /v1/messages/count_tokens (the last one accepts the field but ignores it, so you can count tokens on the exact body you are about to send). The Message Batches endpoint does not currently support customer billing: if one request in a batch sets bill_to: "user_profile", that request errors out with invalid_request_error, nobody is billed for it, and the batch's other requests run as usual. One more rule: a Messages request billed to the customer cannot use billable code execution features, or the API returns a 400. Calls must also carry the user-profiles-billing-2026-10-01 beta header. All of these are beta-stage limits, and Anthropic says endpoints, parameters, and behavior can change during the beta — leave room for that before using it in production.

Three ways to pay, side by side: API keys, platform-funded, linked billing

Putting the three ways to pay side by side makes linked billing's benefits and costs clearer. The first is the API key model: the user pastes their key into a third-party tool, and the tool calls Anthropic's API with it. The benefit is clear billing ownership — the money spent is the user's own. The drawback is that once a key leaks, someone else can spend in the user's name without limit, and the key has to live on a third-party server, with rotation and revocation both painful. Many security teams simply do not allow employees to do this, and that is the core problem linked billing sets out to solve: the first line of Anthropic's announcement is "without sharing an API key", and the security risk of the key model is the whole reason the feature exists.

The second is the platform-funded model: the platform folds the model cost into its subscription price and the user just uses it. The benefit is the simplest possible user experience — nothing to set up. The drawback is that the platform carries the cost risk on its own. Claude usage is hard to predict: one heavy user can burn dozens of times more tokens in a month than a light user, so the platform either prices high and scares off light users, or prices low and gets eaten alive by heavy ones. Many small tools get stuck exactly here: it is not that they do not want to offer Claude, it is that they cannot make the arithmetic work. Linked billing splits that arithmetic apart: the model cost goes on the user's Anthropic bill, and the platform only charges for its own service — pricing becomes clean overnight.

The third is linked billing itself. Its benefit is the combination of the first two: users do not hand over a key, platforms do not front the cost, and the payer can switch per request. Its costs are twofold: the compliance boundary already covered — requests run in the platform's organization, so the customer's data terms do not follow — and the integration effort, since a platform must go through a full registration flow of an OAuth app, a service account, and workload identity federation, plus handle two kinds of refusal (403 and 402) and the gaps around the batches endpoint and code execution features. For a large platform with a mature billing system, that work is worth doing; for a just-launched small tool, the development cost has to be weighed first. Anthropic clearly knows this, which is why it hosts the approval flow and the billing itself — platforms do not have to build metering or invoicing — and that is what "launch faster" in the announcement actually means.

Who benefits first: from coding tools to team collaboration

Judging by the first nineteen platforms, two kinds of product benefit first. The first is coding tools: Factory, Kilo Code, Warp. What they share is that users make heavy Claude calls inside them, and the volume of calls tracks directly with the user's real work. Until now these tools either folded the model cost into the subscription or asked users to paste a key; now users can pay directly through their own Anthropic organization, and the tool's pricing can cover only the service. For people who write code in these tools every day, the change is concrete: no more managing a key pasted everywhere, and the bill is visible in Anthropic's Console, split by platform, showing what each tool spent and on which model.

The second is team collaboration and automation tools: Linear, Asana, Tines, Gumloop, Railway, Ramp. Claude calls inside these tools are usually team behavior: one ticket triggers a summary, one approval flow triggers a classification, one deployment triggers a check. Until now those calls were billed either to the platform or to one person's key, and splitting the cost was awkward. With linked billing, the cost can go straight onto the team's Claude organization bill, the team admin sets a monthly cap, and when the cap is hit it stops — nobody ends up fronting the whole team's bill without knowing. For a team admin, that cap is the most important control: it turns the anxiety of "will this AI tool spend out of control" into a number you can set, change any time, and cancel any time.

Developers who integrate it have real work to do

For platform developers, integrating linked billing is more than adding one field — there is a registration flow to complete first. Create an OAuth app in the Claude Console and select the "Bill and link to organizations" permission; Anthropic issues no client secret, and instead verifies your back end through workload identity federation. Your back end also needs an identity of its own: create a service account and a federation rule, and that service account becomes the OAuth app's only operator service account — one service account can be the operator of only one OAuth app. Then there is the one-to-one rule for user profiles: one customer's Claude organization can be linked to only one of your user profiles, and one user profile can be linked to only one Claude organization. Before linking your first customer, settle which profile stands for which customer, or things will get tangled later.

The approval flow itself is hosted by Anthropic and follows fixed steps: the customer clicks the button, goes to auth.claude.com, picks an organization, reads how billing works, sets a monthly spend limit, and clicks "Link and authorize". Two time limits are worth remembering: the approval page's request_uri is valid for 600 seconds, and the one-time authorization code that comes back in the callback expires after 5 minutes and can be used only once. Your back end must exchange it immediately, together with the PKCE code verifier, for a connection ID and access token, and store them. Anthropic also flags a pitfall that is easy to hit: the approval page only shows organizations belonging to the Claude account signed in to that browser, so if the customer is signed in to the wrong account, the wrong organization gets linked — say so clearly in your interface and ask customers to check the account shown on the approval page first.

The nineteen launch platforms, as a list, already say who Anthropic is aiming at: Asana, ClickHouse, Factory, Gumloop, Harness, Higgsfield AI, Kilo Code, Linear, Major, Mintlify, Omni, OpenArt, Niteshift, Railway, Ramp, Runlayer, Tessera, Tines, and Warp. It mixes collaboration tools such as Linear and Asana, coding tools such as Factory, Kilo Code, and Warp, and infrastructure and finance tools such as Railway and Ramp. Anthropic also says customers can find platforms that offer linked billing on the Claude Marketplace. The list will keep growing, but its first batch is already clear: cover first the places where users are already making heavy Claude calls.

For people doing vibe coding, there is a third group of beneficiaries: small apps vibe-coded around Claude. Until now, the hardest question for such an app was "who pays for the model": cover it yourself and risk runaway usage, or ask users to paste a key and watch them refuse. Now there is an official answer that needs no key: let users link their own Claude organization and charge only for the service — or even offer it free and earn goodwill. The catch is that this only works if your users already have a Claude Platform organization. For developer-facing tools that premise mostly holds; for consumer apps whose users may not even know what a Claude organization is, linked billing does not help. Work out who your users are before deciding whether this road is open to you.

Our judgment: good news for anyone building a Claude wrapper — but do not rush it into production

What this launch really changes is not an interface but the viability of a business model. Until now, building an app wrapped around Claude meant either absorbing the model cost yourself or asking users to paste a key. The first turns pricing into a gamble: when usage rises, margins get eaten. The second discounts both user experience and security. Linked billing offers a third way: users pay with their own Anthropic bill, and the platform charges only for its own service. That matters most for vibe coding — many vibe-coded apps are, at their core, a Claude call with an interface on top, and the hardest question they faced was always "who pays for the model". Now there is an official answer that needs no key.

But two cautions come first. One: it is still a beta, and endpoints, parameters, and behavior can change — the documentation says so explicitly. Integrating now means being ready to change with it; do not treat it as a stable production interface yet. Two: think through the compliance boundary first. Requests run in the platform's organization, so the customer's data terms do not follow; scenarios that need HIPAA, FERPA, or customer-managed encryption keys should not be linked now. And there are the eligibility limits: customers billed through a cloud marketplace, customers with only a subscription plan, and customers hoping to pay with the credits bundled in their plan cannot use it yet — check in Console which category your customers fall into before deciding to adopt it.

One habit is worth building after you link: check each platform's spend in Console once a week. Anthropic's Console shows spend split by linked platform, so you can see at a glance which tool burned the most this month and which cap is about to run out. The value of that habit is that linked billing makes "who pays" transparent, but transparency is not control — only looking at the bill regularly makes the cap actually work. If one platform's usage suddenly doubles, do not raise the cap first; look at where the extra calls came from. It may be a new feature with a high-usage mode switched on by default, or an automation loop calling in circles. Find the cause before raising the limit — it is cheaper than just paying more.

For individual developers, the most practical use runs the other way around: if you already use Factory, Kilo Code, Warp, Linear, or the other supported tools, you may never need to paste an API key again. Link once, set a monthly cap, and the bill runs through your own Anthropic organization, with every dollar visible in Console. Set that cap conservatively: linked billing will not quietly front the money when you run out — it will refuse the request outright. Set the cap too low and your tool stops at the worst moment; set it too high and the bill can run away. Start with a small cap, run it for two weeks, see the real usage, and then raise it slowly.

Views 6Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
A developer types at a dual-monitor workstation; the left screen shows a dark code editor, the right a testing dashboard with green checkmarks and bar charts
News
Did Your Agent Get Better or Worse? AWS Open-Sources UAEF to Measure Agents from Every Framework with One Ruler

On 11 November 2026, AWS open-sourced UAEF: adapters turn traces from LangGraph, LangChain, Strands and Bedrock AgentCore into one format, then 36 metrics across 7 dimensions score them, flagging regressions past 5%. The judge model is hardened too — scores must cite verbatim evidence, and never decide safety.

Developer WorkflowProduct LaunchTesting & Quality
Against a black background, a conductor seen from behind raises a baton, with swirling trails of light circling beneath him like an orchestra being conducted
News
One API for 45 Models, and It Switches Providers When One Fails: Perplexity's Orchestrator Wants to Be the Switchboard

On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.

Developer WorkflowProduct LaunchModel Updates