Control localization cost without hiding failures

Avoid unnecessary work while keeping the usable output, failed attempts, and remaining language requirements visible.

Localization is more efficient when it produces the required usable content with less unnecessary work. A lower total caused by missing languages or failed targets is a different result. Compare spending with what the team can actually use.

Begin with required source revisions and locales. In your operating records, distinguish accepted current output, stale output, absent output, and unresolved work. Do not assume one dashboard reconstructs all of those states for every caller.

Remove repeated work at its source #

Inspect source patterns and target paths. Generated translations must not be discovered as new source content. In a monorepo, confirm which project configuration the runner's working directory selects.

Keep the previous source and reviewed target baselines required by the format. Deleting them to resolve a conflict can make the next request much broader. Investigate with the state preserved first.

Translation reuse, context retrieval, and prompt caching avoid or reuse different work. The first can preserve target text; the second supplies knowledge for generation; the third concerns provider input processing. None means all surrounding operations are free.

Select the update that is actually needed #

lingo push --estimate plans the change against the recorded baseline without translating or creating a job. It uploads files the server lacks for planning, so it is not an offline command. It reports conditional cost for follow-up corrections and spend billed by your own AI provider separately from the planned total. The amount is an estimate rather than a guaranteed final price.

Use ordinary delta translation for changed source, backfill for missing target files, and supported key scope for a selected correction. Finish ordinary source updates before adding a language through backfill, so advancing the baseline does not hide stale existing targets. Force requests new translation deliberately; it is not a routine recovery flag. The Push reference gives exact behavior and format limits.

If one catalog message changed while German and French targets are already accepted, preserve those baselines and translate the intended change. Inspect the diff for unrelated changes. That establishes bounded work and retained output; a savings claim still needs comparable cost evidence.

Compare model costs with repair and coverage #

A provider rate is one component. Inspect actual model and fallback use, output length, failed attempts, repeated work, and reviewer corrections. Compare translation models explains how to keep the source and requirements comparable.

Retain missing cost fields as missing. A chart based on recorded successful requests cannot reveal every unrecorded failure or skipped requirement. Keep provider tokens, provider cost, localization charges, and other account costs in their own units.

Treat a budget limit as a decision about unfinished work #

Understand billing and credit controls before a broad request. If a credit limit interrupts it, keep the run identity and failure, then decide whether to fund, defer, or narrow the remaining work. Repeated submission and silently dropping a locale conceal the outcome rather than improving it.

If lower-priority content is deferred, record the changed requirement and its owner. Preserve high-risk release conditions. Publication records should distinguish an intentional deferral from a failed attempt that nobody accepted.

Report useful output beside known spending #

For the period, report selected workload, usable completed outputs, failures, unresolved coverage, and known cost. Separate generation from measurement and state what the records exclude. Account for changed languages, source volume, model configuration, and review policy before comparing periods.

The resulting decision should explain what work was avoided, why it was unnecessary, and what remains required. That explanation makes a smaller cost meaningful.