What your AI tools cost you, team by team, from the Anthropic Admin API. This is a demo: one file holding a frozen sample, with no server and no API key. Every screen is real Splitmeter code — only the data is invented. Budget edits and key changes live in this browser tab and are forgotten on reload.
Days are UTC. Past days use billed cost from Anthropic's cost report, split by each key's share of tokens (by list-price cost where fast and standard mode share one billed line). Today is still in progress and uses estimated cost from the price table. Usage with no known price is counted in tokens but not in cost.
| Engineer | Keys | Cost | Tokens | % of org | This month vs budget | Last 30 days |
|---|
Click a row for details. Request counts are not shown: the Admin API reports tokens, not requests. / search · j/k move · Enter open · Esc clear · r refresh · b budgets
| Team | Other tools | Where it goes | Claude API | All tools |
|---|
| Person | Plan | Est. cost | Sessions | Lines +/− | Commits | PRs | Tool accepts |
|---|
Can't be seen in usage data: oversized tool definitions, full chat history re-sent each call, retry loops (except as a spend spike). Those need per-request data. See Reality check.
An illustrative case: at 13:20 UTC an agent on one engineer's key got stuck retrying and spent about $38 an hour until they noticed at 23:30. (In mock mode, Priya's key has this spike yesterday.) No guardrail can stop this instantly: usage shows up in the Admin API about 5 minutes after requests finish, and the only lever is turning the whole key off.
Checked against Anthropic's Admin API docs on Sep 22, 2026, and re-checked while building the working version. Rows marked "updated" changed after that re-check.
| Feature | Verdict | Why |
|---|
Hard caps alone aren't a product: the Claude Console already lets you set monthly spend limits and alerts per workspace. That only works in the Console, per workspace (not per key), with at most 100 workspaces. Splitmeter's value is the layer on top: spend per engineer across keys, waste detection with dollar amounts, and fast automatic responses to runaway loops.
The biggest gap is coverage. Cursor and GitHub Copilot bill separately, so a team using several tools needs one integration per vendor. Starting with Claude only is realistic; covering every tool is the long-term reason someone pays.