If your company uses Claude through the API, one question gets harder every month: who is actually spending the money? As of October 29, 2026, Anthropic's Usage and Cost Analytics API is in beta to answer exactly that, with breakdowns by model, workspace, API key and the user or service account behind each request. (Sources: Developers Digest, Anthropic documentation)
The trade-off is speed. Anthropic's older reporting API updates in minutes. The new one updates in hours. That difference decides whether it is a budgeting tool or an alarm system, and it is the detail most announcements skip. Below is what the new API does, what changed from the legacy version, and who actually needs it.
Quick summary: Anthropic's Usage and Cost Analytics API is in beta (beta header analytics-usage-cost-2026-08-24) with two endpoints, one for tokens and requests and one for cost in USD. It groups by model, workspace, API key, service tier and the user or service account behind each request, and covers per-user Claude Code cost. Data usually appears within 4 hours and can take up to 24. The older Usage and Cost API, which updates in about 5 minutes, remains supported but is now labeled legacy. It requires organization-level admin access and is not for individual accounts.
What the new API does
There are two endpoints: GET /v1/organizations/analytics/usage reports token counts and request counts, and GET /v1/organizations/analytics/cost reports cost in USD. Anthropic's documentation says to use it to attribute spend by workspace, API key, model, service tier or the identity behind each request, to rank the top consumers, and to build daily, hourly or per-minute time series for finance reports. (Source: Anthropic documentation)
Developers Digest highlights actor attribution, meaning the API can identify who a request is charged to, and per-user Claude Code cost via products[]=claude_code and group_by[]=actor, which the legacy documentation also points to. It is also available on Claude Platform on AWS and Claude Models on Google Cloud (Anthropic managed). (Sources: Developers Digest, legacy documentation)
The limits to know before you build
Per Developers Digest, summarizing the docs: amounts are decimal strings in cents with six decimal places, so 41280.000000 means $412.80. Data starts January 1, 2026 and covers the last 365 days. A request can span at most 31 days, or 24 hours with one-minute buckets. Up to five dimensions can be grouped and pages hold 1 to 1,000 rows. Values can be revised for about seven days after a month ends. Misspelled model names or unknown IDs match nothing and return no error, which can quietly make a dashboard look cheaper than reality. (Source: Developers Digest)
| Feature | Legacy Usage and Cost API | Usage and Cost Analytics API (beta) |
|---|---|---|
| Endpoints | /v1/organizations/usage_report/messages and /v1/organizations/cost_report | /v1/organizations/analytics/usage and /analytics/cost |
| Data freshness | Typically within 5 minutes | Usually within 4 hours, up to 24 hours |
| Status | Legacy, still supported | Beta; recommended for new integrations |
| Attribution | Model, workspace, API key, service tier and more | Adds the user or service account behind a request, up to 5 grouped dimensions |
| Credentials | Admin API key or admin-scoped credential | Console: Admin API key or org:admin OAuth; Enterprise: key with read:analytics scope |
Anthropic's own legacy documentation says the older endpoints remain supported and recommends the new API for new integrations, so nothing forces a migration. The legacy docs also suggest polling no more than once per minute for sustained use. (Source: legacy documentation)
Which API fits which job
Because of the lag, the two APIs suit different jobs. The legacy API's five-minute freshness makes it better for near-real-time spend alerts. The new API's richer attribution and longer history suit month-end chargebacks, finding the keys that spend the most, and per-user reporting. That recommendation is my reading of the documented numbers, not Anthropic's wording. Developers Digest also advises against real-time spend alarms on the new endpoints. (Source: Developers Digest)
Who actually needs it
Anthropic's docs state that the Admin API is unavailable for individual accounts; you need an organization set up in the Claude Console. So if you use Claude through a personal subscription, this is not for you. It matters for engineering leads, finance teams and anyone allocating AI spend across teams. (Source: Anthropic documentation)
- Use the new API for monthly reports, chargebacks and top-spender rankings.
- Keep the legacy API (or a vendor dashboard) for same-day alerts, since the new data can lag by up to 24 hours.
- Double-check filters, because wrong IDs return empty results instead of errors.
- Plan for revisions: monthly values can change for about 7 days after month end.
Before you wire up a dashboard, pull one known day from both APIs and compare. If the totals differ after the new data has settled, you have a filter or grouping mistake to fix before anyone sees the chart.
01When did the Usage and Cost Analytics API launch?
Anthropic's docs call it a beta but give no launch date. Developers Digest covered it on October 29, 2026, and the beta header carries the version date 2026-08-24, so it may predate that coverage. Requests need the beta header analytics-usage-cost-2026-08-24, which the SDK sends for you.
02How fresh is the data?
Usually within 4 hours and at most 24 hours, versus about 5 minutes for the legacy API.
03Can individuals use it?
No. Anthropic says the Admin API is unavailable for individual accounts; an organization in the Claude Console is required.
04Is the old API going away?
Anthropic labels it legacy but says it remains supported.
05Can it show cost per user for Claude Code?
Yes. Anthropic's docs describe calling the cost endpoint with products[]=claude_code and group_by[]=actor.
If you are watching your AI spend, comparing models on your own tasks is the easiest way to find the best value. LumiChats offers 40+ premium AI models side by side in one subscription.
