Costs

Explain recorded translation and measurement spend without counting overlapping work or treating missing data as free usage.

A cost investigation begins with the work being priced. Generating a translation, reviewing it, and judging it against a criterion are separate operations. Their records can also overlap: one model request can serve several documents and appear in several stage views.

Use request, stage, and measurement evidence to explain individual operations. Use account usage and Billing to reconcile the account period and payment controls.

Identify the accounting unit #

RecordRead with the amount
Translation requestActual model, status, tokens, fallback, and whether it represents an attempt or the consumed output.
Pipeline stageLinked model attempts and any external review order, including its provider and recorded fee.
Measurement resultJudge or algorithm identity, answer usage, and recorded cost.
Account usageBilling period, organization, usage category, and charges or credits.

Count each operation once. Linking several evaluator answers to one request does not multiply its generation cost. Likewise, adding request costs to a job total requires knowing which requests the job amount already includes.

Read the full model attempt #

The request view records inputTokens, outputTokens, and providerCostUsd when the provider supplied them. These figures belong to the whole model request. A batch can serve more than one document or target, so deduplicate request IDs when comparing stages or locales.

The provider amount is not a job invoice. Include failed attempts and successful fallbacks when explaining generation spend, then reconcile platform usage and any external review separately. Prompt caching can affect provider work, but repeated input does not guarantee a cache hit, discount, or latency reduction.

Distinguish absent cost from zero #

Some records do not carry complete accounting evidence. Crowdin events supply no generation usage, and job-derived events can also lack it. Those event fields cannot establish a total generation spend. A dash in the model-request view means the corresponding tokens or provider cost were not recorded; it does not mean zero.

LLM measurement cost can be stored as zero when the provider supplied no cost. The current result response cannot distinguish that from known zero spend. Use independent accounting evidence to establish cost-data coverage, rather than counting every zero as a free operation or a missing amount.

Edit rate and edit distance run without an LLM, so their model spend is zero. Successful measurement runs still have usage accounting. Condition-skipped outcomes, retries, and failures follow different paths; summing visible answer costs does not reconstruct the invoice or every attempted evaluation.

Keep external review fees with the review record #

Open the human-review stage to inspect its external order, provider, and recorded fee. Distinguish an order billed through Lingo.dev from one using your organization’s provider connection. Fees can remain after a team takes over or does not accept a late vendor result. Payment for a review does not establish which text became the accepted output. Inbox reviews covers that decision.

Compare cost with the result obtained #

For a model comparison, keep content, locale pairs, requirements, and time conditions comparable. Include retries, fallback, and additional stages. A cheap successful sample cannot establish a low cost per accepted translation across the workflow.

For measurement yield, a useful calculation is recorded judge cost for the selected completed-answer cohort divided by its determinate-answer count. Include completed null-answer evaluations in that cost numerator, because they consumed work without giving a determinate judgment. Label the amount as recorded answer cost: failed or invalid attempts and retries may be absent.

Show completion and determinate coverage beside that calculation, with the numerator's scope and independently established cost gaps. Control localization cost without hiding failures develops the broader comparison of spend, usable output, and unfinished work.