Skip to content

Usage & costs ​

The Usage & costs tab summarises billed model API calls this project has made — including retries that were later discarded — with token counts and a single USD cost.

The same aggregates are available on the command line as ai-i18n-tools usage.

Use it to answer: how many calls did we make, which models and operations spent tokens, and what did it cost?

What is recorded ​

Each detail row is one HTTP completion that returned usage (even if the translation was later rejected for parse/script/quality and the next fallback model was tried). Transport failures that never produced a billed body are not logged.

After a command that called the API finishes, rows older than seven full UTC calendar days (from 00:00 UTC of today minus 7 days) are rolled into monthly totals (api_totals) and removed from the detail log. The last week stays as individual api_calls rows. Summary tables combine both sources for the selected time range.

Cost reporting ​

Each call, summary card, and table shows one USD cost:

  1. The provider's usage.cost when the response included it (OpenRouter and Cheaper Inference today). A reported cost, including 0, is never replaced.
  2. Otherwise the amount from providers.<name>.modelPricing or the provider-wide pricing default, applied to that call's input and output tokens.
  3. Otherwise the models.dev list price for that model id (input and output USD per 1M tokens), from the same cache suggest-models maintains. suggest-models --refresh updates that cache. A translation or this report does not download languagebench or Arena just to price a call.

New calls store that amount on the api_calls row. Older rows that were stored without a cost are priced the same way when the report is opened, then added into the same Cost figure — they are not shown as a second column. If none of the three sources apply, the cell is —, never $0.00. Changing the configured rates or a later models.dev refresh does not rewrite rows that already have a stored cost. Monthly rollups still keep a count of calls that stored a cost (ncost_acc / ncost_dis) so $0.00 stays distinct from “unknown” after compaction.

Filters ​

Filter by time window, provider, model, operation (translate-docs, translate-ui, translate-json, translate-svg, proofread-ui, bench-models), locale, and outcome (accepted vs discarded).

Time windows:

  • Short ranges (Last 30 minutes through Last 30 days) use rolling durations. Usage over time shows one row per UTC calendar day that still has detail in that range.
  • Last 2 months / Last 3 months start at 00:00 UTC on the first day of the current calendar month minus 1 / 2 months. Usage over time shows the retained daily rows plus one row per month.
  • All time includes every monthly total and the retained daily rows.

To drop old usage, choose a window under Delete entries older than (> 1 month, > 2 months, > 3 months, > 6 months, > 1 year, or all data (clear)) and then Delete data. The menu starts at -, which leaves Delete data disabled until a window is chosen. The same windows are available as ai-i18n-tools usage --clear [--older-than 1mo|2mo|3mo|6mo|1y|all]. Calendar cutoffs keep the current month (1mo) or the current month plus the preceding months.

Command line ​

bash
ai-i18n-tools usage
# ai-i18n-tools usage --since 7d --operation translate-docs
# ai-i18n-tools usage --since 2mo
# ai-i18n-tools usage --clear --older-than 3mo --dry-run

Released under the MIT License.