Anthropic Spend Review
Overview
A client's Claude API bill is driven by tokens, split across workspaces, models, and keys. This skill turns the Admin API's cost and usage reports into the numbers an MSP needs: last month's total, who or what drove it, how it compares to the month before, and what to charge back to whom.
Anti-triggers
- Claude Code seat value (sessions, lines, PRs per developer) — use claude-code-adoption.
- Whether a high-spend key should exist — use key-and-audit-review (admin-gated).
Which tool for which question
| Question | Tool | Notes |
|---|---|---|
| What did the org spend per day, per workspace? | anthropic_admin_cost_report |
group_by: ["workspace_id"] or ["description"]; daily buckets only, 31 per page |
| Which models, keys, or workspaces used the tokens? | anthropic_admin_usage_report_messages |
group_by on model, api_key_id, workspace_id, service_tier and more |
| Spend per person, product, model, or key? | anthropic_admin_analytics_cost |
Beta; group_by includes actor, api_key_id, product, cost_type; max 31 days per call |
| Tokens and requests per person or team? | anthropic_admin_analytics_usage |
Beta; same limits; order_by total_tokens or requests |
| How much headroom before 429s? | anthropic_admin_rate_limits_list, anthropic_admin_workspace_rate_limits_list |
Requests and tokens per minute by model group |
Units
Every cost figure here is in cents, as a decimal string. The cost
report's currency: "USD" names the currency, not the unit: "123.45" is
$1.23. Divide by 100 before you write a dollar amount, and sum the strings
as decimals rather than floats when the total goes on an invoice.
Workflows
Monthly spend review
anthropic_admin_organization_get- confirm the org.anthropic_admin_cost_reportfor the month:starting_atthe 1st at00:00:00Z,ending_atthe 1st of the next month (exclusive),limit: 31,group_by: ["workspace_id"]. Pagenext_pageifhas_more.- Repeat for the prior month. Report total, change, and the workspaces that moved most.
- For the top movers,
anthropic_admin_analytics_costover the same range withgroup_by: ["model"]or["actor"]andorder_by: "amount"to name what drove the change.
Chargeback by workspace (workspace-per-client MSPs)
anthropic_admin_workspaces_list(addinclude_default: true) to mapworkspace_idto client names.anthropic_admin_cost_reportgrouped byworkspace_idfor the billing period.- Sum per workspace, convert cents to dollars once at the end, and list
spend with
workspace_id: null(the default workspace) on its own line. Unattributed spend is a finding: keys in the default workspace should move to a client workspace.
Spike investigation
anthropic_admin_cost_reportdaily for the last 30 days to find the day.anthropic_admin_usage_report_messagesfor that day withbucket_width: "1h"andgroup_by: ["api_key_id", "model"]to find the hour, key, and model.anthropic_admin_analytics_usagefor the same window grouped byactorto name the person or key.- If the key looks wrong, hand off to key-and-audit-review.
Rate-limit headroom
anthropic_admin_rate_limits_list, optionallymodel, for org limits.anthropic_admin_workspace_rate_limits_listwithinclude_inherited: truefor a workspace about to take more load.- Compare against peak per-minute usage from
anthropic_admin_usage_report_messageswithbucket_width: "1m"over the busiest hour. Report the gap, not just the limit.
Gotchas
anthropic_admin_cost_reportcannot group by API key or person. Useanthropic_admin_analytics_costfor that.- Analytics are capped at 31 days per call and 365 days back; a quarter is three calls.
- The cost report and the Console invoice can differ by credits, taxes, and plan fees that are not token usage. Call the figure "API usage cost", not "the invoice".
ending_atis exclusive. A month ending at...-31T00:00:00Zdrops the last day.
Related Skills
- api-patterns - pagination, units, errors
- claude-code-adoption - Claude Code usage per developer