Token Usage Visibility & Attribution: Know What Every Agent Spends
Break every call down by access key, key group, and model — input, output, and cache-hit tokens metered separately at 1-minute granularity, so "who is burning money" stops being a guess.
What the community actually complains about
A classic agent-era bind: one API key shared across a dozen tools, agents, and teammates, and the month-end invoice is a single number. Ask "which agent is burning money" or "which model ate 80% of spend" and you're stuck exporting raw logs and stitching them together by hand.
Shared keys make usage and cost un-attributable, so budget reviews and internal billing never happen.
How GateLLM does it
GateLLM lands every call's usage and cost on decomposable dimensions: the statistics page groups by model, access key, and key group, meters input / output / cache-hit tokens separately, and persists at 1-minute granularity.
The billing page splits monthly cost by access key and expands to per-model detail; Excel export produces 6 sheets including access key × model.
5 levers that actually land
- Group by model / access_key / access_key_group, expandable to detail
- Input, output, and cache-hit tokens metered separately, with latency and success rate
- 1-minute persistence; 24h / 7d / 30d / custom ranges with auto-grained time series
- Monthly cost broken down by access key, expandable to per-model, with totals
- Excel export with 6 sheets: model / access key / key group / key × model / day / week