PRACTICAL GUIDE

AI token estimates versus billed cost

Start from what was billed

Headline input/output token rates may not capture caching, reasoning, tools or the provider actually used. OpenRouter's billing explanation describes these dimensions. Use recorded billed amounts when reviewing actual spend.

Prepare the CSV

project,provider,model,requests,input_tokens,output_tokens,billed_usd
Support,Example provider,Fast model,1000,200000,50000,4.50

This is our normalized format, not a promise to accept every provider export unchanged. Use either request rows or disjoint aggregates. Combining both counts the same spend twice. Convert amounts to USD outside the tool using a documented method before import; do not relabel another currency as USD.

Interpret the monthly pace

If the file contains $140 for 14 complete days in a 30-day month, the linear pace is $300. This does not predict growth, incident retries or future model mix. A per-customer average is a rough allocation: uneven customer usage needs a customer-level breakdown.

Where this tool fits

This is an offline review of a file. It does not enforce budgets or monitor live traffic. LiteLLM spend tracking, Langfuse and Helicone already offer broader tracking options. Use their official pages to evaluate current limits and pricing.

Before changing models

Lower spend is useful only if the task still succeeds. Compare cost per accepted task, retries and latency on your own evaluation set before changing production routing. This tool does not measure model quality.