Budgeting 100,000 Words Across Five Languages

A translation backlog of 100,000 English words is not a single token estimate. When the same material ships in Spanish, French, German, Japanese and Arabic, the source travels through the model repeatedly, while each target produces a different output length. For teams running sustained workloads, the useful estimate separates source tokens, instructions, translated output, caching and rework.

01 · Step 1: Define the five-language delivery scope

Start with an inventory, not a word-to-token shortcut. Record the source volume, target languages, document types and required deliverables. Translating support articles alone is a different workload from translating those articles, generating localized summaries and running a second review pass.

For this walkthrough, assume a hypothetical project containing 100,000 English source words, translated once into each of five languages. Summaries and independent quality review are outside the initial estimate. All token counts below are illustrative planning assumptions, not tokenizer benchmarks or measured language averages.

02 · Step 2: Measure representative source and target samples

Select samples that reflect the actual backlog: ordinary paragraphs, product terminology, lists, placeholders and markup. Tokenize the complete request and response with the selected model's tokenizer where available, then reconcile those counts against API usage records.

Run the same samples for all five targets. Do not assume that equal word counts produce equal token counts, or that Japanese and Arabic follow a fixed multiplier relative to English. The model, script, terminology and formatting all affect the result. If evaluating Claude Sonnet 5.5 and Gemini 2.5 Flash, build separate estimates rather than transferring one model's ratios to the other.

03 · Step 3: Count repeated input across 500 requests

Suppose the project is divided into 100 chunks, each containing roughly 1,000 source words. Sending every chunk separately for each target creates 500 translation requests.

For the worked example, assume each source chunk occupies 1,350 tokens, and each request adds 200 tokens of instructions, glossary entries and formatting rules. That produces 1,550 input tokens per request.

• › Source content across five targets: 1,350 multiplied by 500 equals 675,000 input tokens.
• › Request overhead: 200 multiplied by 500 equals 100,000 input tokens.
• › Total input before caching or retries: 775,000 tokens.

Include any conversation history or contextual excerpts actually sent. Smaller chunks may improve recovery from failures, but they also repeat instructions more often.

04 · Step 4: Build a separate output allowance per language

Use pilot results to replace the following hypothetical output allowances. These figures demonstrate the calculation; they do not predict the relative token efficiency of the five languages.

• › Spanish: 140,000 output tokens for the complete source collection.
• › French: 150,000 output tokens.
• › German: 160,000 output tokens.
• › Japanese: 190,000 output tokens.
• › Arabic: 180,000 output tokens.

The illustrative output total is 820,000 tokens. Combined with input, the baseline is 1,595,000 tokens. Keep the categories separate: adding them is useful for capacity planning, but not sufficient for pricing because input and output can carry different rates.

05 · Step 5: Price caching and rework as separate categories

Calculate spend by multiplying each billable token category by its applicable model rate. Where caching applies, separate ordinary input, cache writes, cache reads and output. Do not treat cached material as free or assume every repeated glossary will qualify for a cache hit.

Add an explicit rework allowance based on pilot failures and review findings. A full-chunk retry repeats more work than a targeted correction. A separate review request also consumes input and output tokens, so budget it as another stage rather than hiding it inside translation overhead.

06 · Step 6: Reconcile the pilot before releasing the backlog

Quanta fits teams that want pay-as-you-go access to multiple models through one balance. Every call is metered on real tokens, with cache writes and cache reads priced separately. Open a Quanta account to run a representative pilot before committing the full backlog.

Track estimated and actual usage by language, model and processing stage. Replace assumptions with observed counts, check output limits, and release work in bounded batches. The result is a translation budget that operations can explain, monitor and revise as the queue grows.