Explore Costs

See LLM spend by operation type — extraction API, manual run, scheduled run — over 7, 30, or 90-day windows.

Cost Explorer answers one question: where did my LLM money go? It pulls from the same usage logs that record every model call you've made — Extract API calls, manual test runs, scheduled test runs, and the pre-populate flow — and lets you slice by time window and operation type.

What you'll see

The Cost Explorer page populated with 30 days of activity: top-line tiles show $1.29 spend, 54 LLM calls, 318.1k tokens used, and $0.02390 average cost per call; a Spend Trends line chart and pie chart break down spend over time and by operation; Spend Attribution panels chart spend by model and by document; a Volume & Reliability panel shows error and retry stats; a long Call Detail table lists every individual call with timestamp, operation type, model deployment, schema, tokens, cost, and latency — Window selector (default: Last 30 days)
Operation selector (default: All operations)
Top-line spend / calls / tokens tiles
Spend Attribution charts
Per-call detail table

The two filters

Window — Last 7 days, Last 30 days, or Last 90 days. Anything older than 90 days isn't accessible here; if you need long-tail analysis, export from the underlying llm-usage-logs (super-admin only).
Operation — All operations (default), or one of:
- Extract API — calls to the /api/extract endpoint from your own applications.
- Manual Test Run — extractions triggered from the Manual Test Run page.
- Scheduled Test Run — extractions triggered by the Scheduled Test Runs cron.
- Pre-populate — extractions kicked off by the ground-truth pre-populate flow (used while filling in Ground Truth Responses).

What's in a populated view

Once you've made some extractions, the page shows:

Spend totals for the selected window, broken down by operation type and by model deployment.
Token totals (input and output separately, since output costs more).
Trend chart over the window, so you can spot a sudden spend spike from a runaway scheduled run or a buggy app integration.

Common questions

"Why is Extract API spend so high this week?" — switch Operation to Extract API, look at the model breakdown. Often a single deployment dominates because your app picked the most expensive one.
"Which scheduled run is most expensive?" — switch Operation to Scheduled Test Run. If one schedule is dominating, consider running it on a smaller test set or a cheaper model.
"Did I get billed for a manual experiment last week?" — set Window to Last 30 days, Operation to Manual Test Run. Yes, you did.

What's not here

Costs for non-LLM operations — OCR, storage, etc. — aren't tracked here.
Costs broken down by individual user. Cost is rolled up at the account level.

Run a Manual Test — the most likely source of a budget spike while you're iterating.
Schedule Recurring Tests — the most likely source of recurring spend; tune cadence + test-set size to match the value you're getting.
Browse Supported LLM Models — to compare per-million pricing before picking a model.