Claude API spend by initiative: using Anthropic's Usage and Cost Admin API
What Anthropic's Usage and Cost APIs and Console exports return and leave out, and how to split a Claude bill by initiative, as of November 8, 2026.
Roiva wrote this, and Roiva reads the APIs it describes. The facts about Anthropic's reports were checked against Anthropic's published docs on November 8, 2026, and each links to the page it came from. Since October 29, 2026, Anthropic calls the Usage and Cost API legacy, still supported, and points new integrations at its beta Usage and Cost Analytics API; both are covered below. Anthropic changes these APIs often, so check the link before you build on a detail.
How do you report Claude API spend per AI initiative?
Give each initiative its own workspace in the Claude Console, then read Anthropic's cost report grouped by workspace. Anthropic's bill has no idea which initiative a request served; a workspace is the one dimension you control that both of Anthropic's cost reports can split by. Anthropic documents workspaces as the way to "separate different projects, environments, or teams while maintaining centralized billing and administration" (workspaces).
A workspace is a container inside a Claude Console organization with its own API keys and spend limits. Its costs show up under its own ID in the cost report.
Which Anthropic report gives you what?
Anthropic reports Claude spend in six places, and which one you need depends on how your company pays for Claude.
| Report | What it gives | Grain | Key or access |
|---|---|---|---|
| Console Usage page | Tokens by model, workspace, API key, month or day, with CSV export | Daily | Console sign-in |
| Console Cost page | Dollars by workspace, model and month, with CSV export | Monthly view | Console sign-in |
Usage API (legacy), /v1/organizations/usage_report/messages |
Tokens by model, workspace, API key, service tier and more | 1 minute, 1 hour or 1 day | Admin API key |
Cost API (legacy), /v1/organizations/cost_report |
Dollars by workspace and by line description (model, cost type, token type) | 1 day only | Admin API key |
Usage and Cost Analytics API (beta), /v1/organizations/analytics/usage and /v1/organizations/analytics/cost |
Tokens and dollars by workspace, API key, model, service tier, product, and the person or service account behind each request | 1 minute, 1 hour or 1 day, or the whole range | Admin API key, or an Analytics API key for Claude Enterprise |
Claude Code Analytics API, /v1/organizations/usage_report/claude_code |
One record per user per day: sessions, lines, commits, accepted edits, estimated cost | 1 day | Admin API key |
Sources: Usage and Cost API, Console reporting, Usage and Cost Analytics API, Claude Code Analytics API.
Claude Enterprise organizations on claude.ai are separate. They don't appear in the Claude Console and carry no Admin API keys, so they read their usage and cost with an Analytics API key instead. Anthropic now recommends the same Usage and Cost Analytics API for them, and calls the Claude Enterprise Analytics API's own cost and usage reports legacy (which API do you need?).
Is Anthropic's Usage and Cost API deprecated?
No. On October 29, 2026, Anthropic made it a legacy API that "remains supported," and recommended the beta Usage and Cost Analytics API for new integrations (release notes). As of November 8, 2026, Anthropic has given no shutdown date.
Every request to the Analytics API must send the
anthropic-beta: analytics-usage-cost-2026-08-24 header, and Anthropic's
migration guide
lists what changes. The differences that matter for splitting a bill:
| Legacy Usage and Cost API | Usage and Cost Analytics API (beta) | |
|---|---|---|
| Cost grouped by | Workspace and line description | Up to five of workspace, API key, model, service tier, product, actor and more, plus cost type and token type |
| Cost by person | No | Yes, by actor (see below) |
| Default workspace | null workspace_id |
Its real workspace ID |
| Freshness | Typically within 5 minutes | Typically within 4 hours, up to 24; figures can change until about 7 days after the month ends |
| Date range | Any | From January 1, 2026, within the last 365 days, at most 31 days a request |
| Claude Platform on AWS | Not available | Available, one AWS region per request |
| Claude Enterprise | Not available | Same endpoints, with an Analytics API key |
What does the Usage and Cost Admin API need?
An Admin API key, which starts sk-ant-admin01- and which only organization
members with the admin role can create. Anthropic also accepts an OAuth token
with the org:admin scope, or a personal or service account key that isn't
scoped to a workspace; "workspace API keys don't work"
(Usage and Cost API).
The Admin API isn't available to individual accounts, so set up an
organization first. For a Claude Console organization, the Analytics API
also takes an Admin API key or another Admin API credential.
What else to know before you build on the legacy endpoints, as Anthropic documents them:
- Money is in cents. "All costs are in USD, reported as decimal strings in lowest units (cents)." The Analytics API's amounts are cents too, with six decimal places.
- The cost report is daily only, up to 31 days a page, grouped by
workspace_idanddescription. Grouping by description adds the model, the cost type (tokens, web search, code execution) and the token type. - The default workspace has no ID. Its usage and costs carry a
nullworkspace_id, so a script that drops nulls drops the default workspace. The Analytics API gives it its real ID instead. - Data is near real time. It "typically appears within 5 minutes," and Anthropic supports polling once a minute. The Analytics API is slower and can revise a month for about a week after it ends.
- Pagination uses
has_moreandnext_page.
Can you see Claude API spend by user?
Yes, with the beta Usage and Cost Analytics API, for requests a person or
service account made. It groups cost by actor: the user, service account or
federated cloud identity behind each request. Personal keys and service
account keys report as the person or service account they belong to. Legacy
workspace API keys aren't linked to anyone, so their requests report the key
itself as the actor
(actors).
The legacy cost report has no user dimension, and in the Console, Anthropic's support article says "currently, it's not possible to break down usage or cost by individual users" (Console reporting).
For Claude Code, Anthropic points to the Analytics API's cost endpoint
grouped by actor and filtered to the claude_code product, and the Claude
Code Analytics API gives per-user, per-day activity with an estimated cost.
For an application your team built on shared keys, log the user, the task
and the engagement or customer in the application itself, and use
workspaces or API keys to split the bill.
What isn't in Anthropic's cost report?
Several kinds of Claude spend never reach the legacy cost report, and a total built only from it will be short by exactly those. The Analytics API closes the gaps for Claude Platform on AWS and Anthropic-managed Google Cloud, but not for Priority Tier:
| Spend | Where it shows instead |
|---|---|
| Priority Tier | Not in the legacy cost endpoint: "Priority Tier costs use a different billing model." The Analytics API shows requests under a Priority Tier commitment at no token cost, and the commitment fee isn't in its data. Track Priority Tier usage through the usage endpoints (Usage and Cost API, Analytics API) |
| Claude on Amazon Bedrock | Your AWS bill; Anthropic lists its Admin and Usage and Cost endpoints among the features Bedrock doesn't support (Claude on Amazon Bedrock) |
| Claude on Google Cloud | Your Google Cloud bill. For the Anthropic-managed offering, the legacy API has only the usage endpoint, and cost shows on the Console Cost page or in the Analytics API, called with Google Cloud credentials (Google Cloud, Analytics API) |
| Claude Team and Enterprise seats | Your invoice from Anthropic, and the claude.ai analytics for Team and Enterprise |
| Claude Platform on AWS | The Console Usage and Cost pages, or the Analytics API, one AWS region per request; the legacy endpoints aren't available there (Analytics API) |
How do you split one Claude bill across initiatives?
Claim what you can by workspace, and split only what's left. A method a finance team can audit:
- One workspace per initiative, created before the initiative goes live, with its own API keys. Spend that lands in the right workspace needs no allocation at all.
- Decide the default workspace's fate. Either keep it empty or treat it as shared. Anthropic doesn't let you set a spend limit on it.
- Split out a shared workspace's keys. Anthropic's beta Usage and Cost Analytics API gives dollars by API key, so a key that serves one initiative can be claimed alone.
- Split the rest by a stated weight, such as requests or seats by team, and write the weight down before the month closes.
- Book the API cost once. The Anthropic invoice in your ERP is the same money as the cost report. Use one as the source for each month, not both.
- Add what the report misses. Priority Tier, Bedrock and Google Cloud spend come from their own bills.
- Watch unit cost, not just spend. Cost per million output tokens, by model and month, says whether a cheaper model or better caching made each answer cheaper, separately from how much the initiative was used.
How does Roiva book Claude API spend to initiatives?
Roiva does the routine above every night, from the same reports.
- It reads the cost and usage reports. Connect Anthropic with an Admin API key and Roiva reads the legacy cost report by workspace and line description, and the legacy usage report by model and workspace, for the last 12 months, then every night after. It never reads a prompt or a response.
- Workspaces become initiatives. When you link the connection to an initiative, Roiva asks how much of the bill is the initiative's: all of it, only some workspaces, each booked whole, or a share of every charge. What no initiative takes stays unattributed rather than being spread silently. A changed share applies from the first of that month.
- An API key can stand alone. An initiative can claim one key and book what Anthropic's beta Analytics API says that key cost. The rest of its workspace books as if the key weren't claimed, and a day whose key figures don't add up to the workspace's bill stays with the workspace.
- The invoice isn't counted twice. In any month the connection books, an Anthropic invoice mapped to the same initiative in QuickBooks, Xero or NetSuite is left out, unless an allocation rule or a locked month keeps the invoice instead.
- Unit cost, with the cause. Roiva tracks cost per million output tokens by model and month. Where an initiative books a saving on that unit cost, Roiva splits the saving into what Anthropic's price changes gave and what your own choices earned: the models you ran, what you cached and how you shaped the work.
- Claude Code and Claude Enterprise. The same key reads Claude Code's daily totals, with user names dropped before anything is saved. Claude Enterprise seats need their own connection with an Analytics API key, which also books usage cost by user group or estimated chat Project. Claude on Bedrock is booked from the AWS connection, by tag or by service, and on Google Cloud from the Google Cloud connection.
Roiva books from Anthropic's cost report, so Priority Tier spend, which isn't in that report, has to reach the initiative as a cost you enter. The Anthropic integration page shows a worked month, and How costs reach an initiative covers every way spend becomes an initiative's cost and where the same dollar can arrive twice.
Book each initiative's Claude spend automatically
Roiva reads Anthropic's cost report every night and books each workspace's spend to the initiative that ran it. A free account has no expiry and connects Anthropic.
Get the next one
New field notes by email as they're published. Or subscribe to the RSS feed.