Tools · BillCheck

Is your AI bill counted right?

A weekly check of a LiteLLM proxy's spend logs. It prices every request again from the public catalogue as it stood that day, and shows the requests priced with an old price, the ones logged at $0 that no budget can see, where the money went, which budgets will run out and when, and where the same work would cost less. It runs in your browser; your logs never leave this page.

1Checked test set3,882 requests sent through a real proxy, each with a known right answer.

2BaselineTrusting the proxy's own number: 42% of requests right, $15.71 off.

3Two approaches comparedRepricing alone (89% right) against repricing plus the proxy's own records (100%).

4Cost and speedAbout 4,000 log rows a second; the example's 4,246 rows in under a second here.

5Error analysisWithout the proxy's records, an old price looked like a discount. Fixed and tested.

6Free to useNo sign-up. Nothing is uploaded anywhere.

Two weeks of a small company's proxy: four teams, nine keys and eight models, sent through a real LiteLLM proxy (version 1.105.1) whose price map was two months old. The traffic is made up and the models were a stand-in, so nothing was paid; every number below is the proxy's own.

Save this query as export.sql. It only reads, and takes the last 14 days and only the columns the check needs, which keeps the file small.

Then run it against the proxy's database:

For budget alerts, export the budget tables too:

Or save the proxy's own API answers as JSON: GET /spend/logs/v2?start_date=...&end_date=...&page=1&page_size=100 with an admin key, one file per page.