Benchmarks
Processing speed
Measured AI translation times from two test campaigns in October 2026, with their workloads, settings and limits.
What this covers
These measurements cover translation only: a Translator run on a job, from start to finish. They do not include file import or export, waiting in a queue, Term Expert, pre-translation, human editing, review or delivery. They describe specific test campaigns, not how long your documents will take.
Document-level translation compared
On 11 October 2026 we compared two ways the Translator can work: segment by segment, and with the whole document in view. TilBridge adopted the document-level approach after this comparison.
- Workload: seven synthetic test documents, including English → Kazakh and German → Russian material. No client documents.
- Environment: a local test system, one process and database per document.
- Model and settings: OpenAI gpt-6-luna with low reasoning. Three repetitions per document and setting; all runs completed.
| Translator setting | Median time per document run | Rows left untranslated (21 runs) | Provider cost per translated row |
|---|---|---|---|
| Segment by segment, no draft | 56.1 s | 34 | $0.000157 |
| Whole document, no draft | 13.3 s | 3 | $0.000082 |
| Whole document, DeepL draft first | 19.8 s | 4 | $0.001076 |
Scroll the table horizontally to see all columns.
Rows left untranslated are rows the Translator could not complete, mostly because of inline tag structure; they remain for a person to translate. A machine-translation draft before the Translator made runs slower and, with DeepL, more expensive.
Runs on real documents
On the same day we ran the Translator on real working documents. Their content is not described here.
- Test 1: a form document, Russian → English, one job of 365 segments; translation memory matched about half of the rows exactly.
- Test 2: three short product-label documents, Russian → Kazakh, 57 segments (about 430 words) translated as one set.
- Environment: a local test system using the same background processing as the app. Three repetitions per setting, low reasoning.
| Test | Model and draft | Completed runs | Median run time | Provider cost per run |
|---|---|---|---|---|
| 1 | gpt-6-luna, no draft | 3 of 3 | 52 s | $0.011 |
| 1 | gemini-3.8-flash, no draft | 3 of 3 | 63 s | $0.074 |
| 1 | gpt-6-luna, DeepL draft first | 3 of 3 | 84 s | $0.074 |
| 2 | gemini-3.8-flash, no draft | 3 of 3 | 49 s | $0.034 |
| 2 | gpt-5.6-terra, no draft | 3 of 3 | 71 s | $0.099 |
| 2 | gemini-3.8-flash, DeepL draft first | 3 of 3 | 77 s | $0.097 |
Scroll the table horizontally to see all columns.
In both tests a DeepL draft made runs slower and more expensive. In Test 2, gpt-6-luna completed only some of its runs because of the documents’ inline tags, so it is not shown; the other models left no rows untranslated.
Limits of these measurements
- The documents are small and few, in several language pairs. They do not represent every document type, size or language pair.
- Times are medians of completed runs on a local test system, not production load or a service-level target. Provider response times vary.
- Costs are provider costs recorded by the app at the time and change with provider pricing.
- These are speed and cost measurements, not a translation-quality ranking.
Planning your project
Use these figures to understand the order of magnitude, then time a representative sample with your own language pairs, models and providers during an evaluation. Allow separately for preparation, human editing, review and delivery checks.