Summary

Token counts and German writing quality need separate tests. Measure your actual texts and review client messages for wording, facts and register.

What you'll learn

  • Treat the corpus result as a dated measurement, not a universal surcharge.
  • Read what a benchmark tests before comparing its scores.
  • Use representative client messages and review the actual output.

The worry is specific and reasonable. Your agent is going to write to your clients in German, and you are afraid it will produce something that reads fine and is subtly, embarrassingly off — the register slightly wrong, the phrasing slightly translated, the sort of thing a German client notices immediately and never mentions.

Treat token use and writing quality as separate decisions. A cheaper text is not necessarily a better message, and fluent German can still contain the wrong amount, deadline or salutation.

The token difference is measurable for a specified corpus. The 63-pair snapshot below has a median 21.5% German premium under o200k_base. That does not set your total agent bill: the actual encoding, input/output prices and surrounding work still matter.[1][8]

Quality needs a different test. A study of translation-like wording can identify a particular linguistic effect. It cannot tell you that every current model writes business German as well as English. Use the findings to choose review questions, then test your own workflow.[2]

Why does German cost more to run than English?

In the August 2026 Yardwork snapshot, German used a median 21.5% more tokens under o200k_base. Tokens are the text units counted for this measurement; they are not words. The same texts give a different ratio under another encoding.[1][7]

The measurement covers 63 published article pairs from the same site, frozen before this article was published. The preparation removes frontmatter, image components, table rows and link syntax, then compares the remaining text. It measures these published versions, not controlled translations of identical wording or a representative sample of all German business writing.[1]

Measureo200k_basecl100k_base
Median German-to-English token ratio1.2151.504
Mean1.2041.473
Range across pairs0.908 – 1.3661.097 – 1.680
Pairs where German used more tokens60 of 6363 of 63
Corpus totals157,740 EN / 189,087 DE tokens—

Two things in that table matter more than the headline.

The encoding matters. The same corpus has a median 50.4% premium under cl100k_base and 21.5% under o200k_base. This comparison does not forecast future tokenizer improvements or establish the encoding of every current model. Ask which encoding your actual service uses.[1]

Three pairs went the other way. German used fewer tokens in three of 63 pairs under o200k_base. The overall tendency in this corpus is clear, but it does not guarantee a higher count for every bilingual text.[1]

Why German output costs more even when the words are shorter

The German versions in this corpus use fewer words at the median and still use more tokens. Compare both measures before estimating the cost of your own correspondence.[1]

Ratio, German ÷ English (median)Value
Words0.907
Characters1.081
Tokens1.215
Tokens per wordEN 1.208 · DE 1.654

The median word ratio is 0.907 and the character ratio is 1.081: about 9% fewer words and 8% more characters in German. The median tokens-per-word values are 1.208 for English and 1.654 for German, a derived 36.9% difference between those medians.[1]

A word count alone can therefore be misleading. A German compound may remain a single word for a reader while the encoding represents it as several pieces. The precise split depends on the encoding and surrounding text.[7]

For a concrete illustration, The data protection officer approves the request has seven words and seven tokens under o200k_base. Der Datenschutzbeauftragte genehmigt die Anfrage has five words and ten tokens. The measured split includes Datenschutz | be | auftrag | te and gene | hm | igt; leading spaces belong to some token pieces.[1][7]

At the same per-token rate within the same billing category, more tokens mean a larger text charge. Compare the actual workflow’s input, output and other charges before turning this corpus ratio into a budget.[8]

The same sentence shown twice as rows of labelled boxes, one box per token: an English row of seven boxes and a German row of ten, with seven of the German boxes highlighted to show two words split into fragments.
Seven words and eight tokens in English; five words and ten tokens in German. Encoding o200k_base.

What gives AI-written German away?

The August 2026 translationese study found a difference between automatic classification and human judgement. For German Gemini 2.0-Flash-001 output, the classifier labelled 60.2% translation-like; the native annotator labelled none of the small annotated subset that way. For Llama 3.3-70B-Instruct, human labels reached 37.5%, including 12.5% ambiguous cases.[2]

The study covers five languages; human annotation covers German and Spanish. The human pool contains 60 sentences per language across configurations, not 60 sentences for every model. Translation-like wording is the target, not factual accuracy, general writing quality or whether a reader can detect AI authorship.[2]

A translation-like style score and a reader’s judgement answer different questions.

For your own client work, review the model you plan to use, with its actual instructions. These particular results do not establish a general German-versus-English quality gap or rank all current models.

Annotator explanations mainly concerned unusual word choices and combinations. That gives a reviewer something concrete to look for: a phrase may be grammatical yet sound unfamiliar in the situation. Keep factual accuracy, tone and task completion on separate review lines.[2]

Try a glossary and approved examples when wording is the problem, then repeat the test. If the message contains a wrong fact or misses a business condition, examine the source information and workflow as well. Do not assume either category is permanently fixed or impossible to improve.

The authors discuss possible explanations involving translated training material and cross-language processing. These are hypotheses about the observed pattern, not measurements of how your chosen model handles your clients’ correspondence.[2]

Use that distinction alongside what AI agents are genuinely bad at: inspect the particular failure rather than judging the whole workflow from fluent wording.

Two horizontal bands: an upper band showing a clearly displaced marker labelled as cost, and a lower band showing two markers close together with a small red highlight on a narrow region at one end.
Measure token use and review writing quality separately; the corpus ratio does not quantify a quality gap.

What can the German benchmarks actually tell you?

The benchmark’s task and construction determine what its score can tell you. OpenGPT-X’s 2024 release used machine-translated questions; EU MMLU’s 2026 release used human translation and revision. Neither method alone scores the emails your agent will send.[4][5]

Read the method before comparing headline scores. Use the benchmark for its stated task, then keep a separate acceptance test for your correspondence.

SourceWhat it can supportWhat it cannot
OpenGPT-X European LLM Leaderboard (2024)[5]Comparison across 21 European languages on a consistent setA claim about native German — ARC, HellaSwag, TruthfulQA, GSM8K and MMLU were machine-translated into those languages with DeepL
EU MMLU (DG Translation, July 2026)[4]Human-reviewed multilingual evaluation, 16 EU languages reported by the announcement, over 1,000 questions, around 250 student translators from 21 universitiesResults for the commercial models you would actually deploy — it is a dataset, not a leaderboard of vendors
LEXam (ICLR 2026)[3]Legal reasoning ability, in each of two languagesA German-versus-English gap. See below
Translationese generation study (COLM 2026)[2]Whether generated German carries the fingerprints of translationA general quality score

Machine translation can introduce its own wording and meaning changes. DG Translation’s stated goal for EU MMLU is to preserve meaning, difficulty and testing value across languages through human translation and revision. That is a method improvement to consider, not a published verdict on the commercial model you intend to deploy.[4][5]

LEXam uses independent English and German questions from different exams, rather than parallel translations. A score difference therefore mixes language with differences in question content. It cannot isolate a German-versus-English language penalty.[3]

The conventions a benchmark never tests

Knowledge and legal-reasoning scores cannot replace a check of how your own messages address the recipient. Specify these conventions and include them in your acceptance review.

Sie or du. German has a formality distinction English does not, and getting it wrong in a client email is a real error with no English equivalent. No knowledge benchmark tests it. It has to be an explicit instruction and a review item — this site's own German content specifies formal Sie throughout as a standing rule, not a preference the model infers.

Salutation and closing conventions. German business correspondence has settled forms. A model that scores well on German factual questions can still open a letter in a way that reads like it was composed in English.

Terminology that should not be translated. Product and tool names stay in English in German business writing — Gmail, Slack, Notion. A model translating them produces text that is grammatically fine and immediately wrong to the reader.

Regional variation. Austrian and Swiss German differ from German German in vocabulary and, for Switzerland, orthography. None of the evidence here measured that, and no claim is made about it.

Clear instructions can help, but they do not guarantee that the model follows every convention. Test the resulting messages and keep reviewing the kinds of mistakes that matter to your clients. The same principle applies to reliability generally.

How to test whether your agent’s German is ready for clients

For a first review, try twenty representative messages and a native German reviewer. Twenty is a proposed starting set for finding problems, not a validated sample size or a guarantee that the work fits an afternoon.

1

Take twenty real messages

Pick twenty German messages you have already sent to clients. Real ones, across the range of situations you handle — not invented test cases.
2

Have the agent write its version

Give it the same context you had and let it draft each one. Use the model and the instructions you would deploy, not a bare chat window.
3

Mark only what you would change before sending

Have a native speaker mark edits they would make before the message goes out. Not preferences — changes they would actually make.
4

Sort the edits into two piles

Word choice and register in one pile; facts, structure and grammar in the other. Use the first pile to test a glossary and examples. Investigate source information, task instructions and model behaviour for the second; repeat both kinds of case after changes.
5

Measure the token difference on your own text

Run the same content through a tokenizer in both languages. Your ratio will differ from ours — 21.5% is one publisher's corpus, and your domain vocabulary will move it.

Record the edits by type and seriousness. Re-test after a change, include new and unusual cases, and require human approval where a wrong message could cause harm. A small first review can uncover problems; it cannot establish the rate of rare failures or prove every client message is ready.

What the token counts do not show

The token measurement concerns a fixed 63-pair publication snapshot. It does not control for equal wording, text length or how either locale was drafted. Keep it as a dated measurement of this corpus, and measure your own traffic before budgeting.[1]

The human annotation in the translationese study is small and uses author-annotators. It complements the automatic analysis rather than providing a universal business-writing verdict. A separate study of translation systems traces translationese to supervised-training bias; that translation task is distinct from generating your client messages.[2][6]

Nothing here addresses translation quality, which is a different task from generation. Nothing here addresses Austrian or Swiss German. And nothing here measured spoken German or transcription, where the limits of what a model can read are a separate question with separate evidence.

Use the corpus results to see why token counts matter, and the research methods to see why a benchmark score is not a client-writing test. The practical decision needs the actual model, the actual correspondence and a review of facts as well as wording.

Frequently asked questions

Do AI models write German as well as English? Test your actual business writing rather than assuming parity. The 2026 translationese study found classifier/human differences in specific Gemini and Llama outputs, with small human annotation sets. That is not a general German-versus-English quality score. Review wording, facts and register separately.[2]

Why does German cost more than English with an AI agent? In the fixed 63-pair Yardwork corpus, German used a median 21.5% more tokens under o200k while using about 9% fewer words. The encoding and the texts determine the ratio. It is not a universal German surcharge or a measure of the whole agent bill.[1][8]

Is there a benchmark that shows how good AI is in German? There are benchmarks for particular tasks. OpenGPT-X used DeepL-translated questions across 21 languages; EU MMLU used human translation/revision for 1000+ questions with nearly 250 students from 21 universities and reported 16 EU languages. Their construction matters, and neither replaces testing your client messages.[4][5]

Does an AI agent use formal Sie or informal du in German? Specify the required register, salutation, sign-off and terminology explicitly, then inspect whether the output follows them. A knowledge score does not answer whether the model will address this client appropriately.

How do I check whether an AI agent's German is good enough for my clients? Try twenty representative past messages as a first review, not a validated sample-size guarantee. Ask a native reviewer to mark necessary changes, separate wording from factual or task errors, and re-test after adjustments. Include unusual cases before granting live authority.

Notes

  1. Yardwork bilingual corpus token measurement. August 24, 2026 snapshot; reproduced October 15, 2026 Accessed October 15, 2026.
  2. An Investigation of Translationese in the Generations of Multilingual Large Language Models. August 18, 2026 Accessed October 15, 2026.
  3. LEXam: Benchmarking Legal Reasoning on 340 Law Exams. April 2, 2026 revision Accessed October 15, 2026.
  4. Towards fair multilingual AI: EU MMLU, a new EU benchmark for LLMs. July 22, 2026 Accessed October 15, 2026.
  5. European LLM Leaderboard. July 12, 2024 Accessed October 15, 2026.
  6. Lost in Literalism: How Supervised Training Shapes Translationese in LLMs. March 6, 2025 Accessed October 15, 2026.
  7. tiktoken. Living repository; tokenizer version 0.14.0 used Accessed October 15, 2026.
  8. API pricing. Living pricing, October 2026 Accessed October 15, 2026.