ModelMeter

THE WORKBOOK / ModelMeter

AI API Cost Tracking Template: Billed Spend by Project

Start from billed cost when reconciling API spending. Token counts alone cannot account for every discount, cache rate, tool charge or billing adjustment.

By ModelMeter project team · Published · Fictional examples, reproducible calculations

Review the sample costs →Download the usage CSV

A 14-day example with three cost buckets

These provider and model labels are fictional. They illustrate the normalized format accepted by ModelMeter, not current provider prices.

Project / modelRequestsBilled USDUSD / 1,000 requests
Support / Fast model12,000$48$4
Research / Reasoning model500$86$172
Support / Backup model800$6$7.50

Total billed spend is $140 over 14 complete days. Research accounts for $86 / $140 = 61.43% of that spend, despite having fewer requests. This identifies a bucket to inspect; it does not establish that the model is wasteful or replaceable.

Monthly pace is an assumption

monthly_pace = billed_spend / complete_days × days_in_month
$140 / 14 × 30 = $300
$300 / 40 active_customers = $7.50 per customer

This calculation assumes the remaining days behave like the observed days. A launch, trial cohort or traffic spike can invalidate it. It is neither a forecast nor your next invoice. An entered $300 monthly budget is exactly matched by this example's pace.

Prepare a comparable export

  1. Keep project, provider, model, requests, input_tokens, output_tokens and billed_usd headers.
  2. Aggregate to non-overlapping rows for the same complete time window.
  3. Reconcile the billed USD column with the provider's invoice or dashboard. Resolve currency conversion and credits outside this tool.
  4. Enter complete days, calendar days in the month and the customer count for your chosen definition.
  5. Review expensive groups before changing routing, quality thresholds or quotas.

What does this report omit?

Costs absent from your CSV remain absent, including infrastructure, storage, retrieval, tool calls and support. Provider-native exports are not automatically mapped. Input and output tokens are retained as context; ModelMeter does not apply a hidden price table to them. Files are processed locally, with no provider key required.