“In der heutigen schnelllebigen Welt ist es nicht nur wichtig, sondern auch entscheidend …” Anyone who uses AI to write German knows this sound: grammatically correct, but somehow translated. Too many anglicisms, a switch from formal “Sie” to informal “du” mid-text, English quotation marks, stock phrases nobody in a German office would ever say.
The obvious question: which language model writes the best German? The honest answer is surprising: there is no test that reliably measures it. The well-known leaderboards test knowledge, reasoning or chat users’ preferences, not the quality of German business texts. That’s why a different model leads depending on the leaderboard.
This guide puts the leaderboards in context, compares the most important models for German texts, shows how to recognize good business German, and gives you the tool that delivers the biggest quality jump: a system prompt for your German house style.
Why AI texts often sound clumsy in German
Three causes explain why German AI texts sound “translated” more often than English ones:
1. German is a minority in training data. Language models learn mostly from web text. In the current Common Crawl, one of the most important public data sources, around 5.8% of pages are German, but 42.6% are English. Manufacturers don’t disclose their exact training mix, but the direction is clear: English patterns shape the models more strongly.
2. Models often “think” in English. A 2025 test by data analyst Malte Landwehr showed that models with an explicit reasoning step frequently process German tasks internally in English. Research on “translationese” in AI texts confirms that this shows in sentence structure and word choice. In a study on Dutch and French, 16% of all flagged phrasings were clearly related to English.
3. German is more expensive to process. Language models split text into tokens. German words, especially compounds like “Gerichtsverfahren” (court proceedings), are split into more pieces than English ones. A study presented at NeurIPS 2023 found around 58% more tokens for the same text in German with the GPT-4 tokenizer. Newer tokenizers narrow the gap; depending on the model, it’s roughly 15 to 60% more today. That means higher cost per text and less room in the context window.
What the leaderboards really measure
Searching for the “best LLM for German” leads to leaderboards. None of them measures what companies actually want to know: whether a model writes natural, stylistically good German business texts.
| Leaderboard | What it measures | Who leads | Limits |
|---|---|---|---|
| Arena, German ranking | Blind user votes: which of two answers to a German prompt is better? | Gemini 4 Argon, followed by a tight group including Claude Opus 5.5 and GPT-5.6 (as of Nov 9, 2026) | Places 2 to 15 are statistically tied within their confidence intervals. Lay chat preferences, not professional text quality |
| i6eal German AI Benchmark | 24 German business tasks including “language quality & style”, scored by AI judges | Claude ahead of GPT and Qwen (run from June 2026, now an older model generation) | AI judging AI; per the operator, differences of a few points are noise |
| EuroEval German | Language understanding: grammaticality, reading comprehension, summarization, knowledge | A mix of open models and Gemini (as of April 2026) | Current top models missing; measures understanding, barely writing |
| Artificial Analysis, German | Knowledge and reasoning questions in German | Gemini and Claude at 93 to 95 points | Nearly saturated, barely separates top models; no statement on style |
What can still be read from them:
- The top models from the major providers are close in German. Anthropic, for example, states that Claude Opus 5.5 reaches 99.1% of its English performance in German on a multilingual knowledge test.
- Grammar varies widely between model classes. In EuroEval’s grammaticality test, top models such as Gemini 3 Pro or Claude Sonnet 4.6 scored above 70, while the German research model Teuken-7B barely registered.
- AI-based ratings should be read with caution. Studies show that AI judges prefer their own texts and tend to overrate non-English texts.
The most important models for German texts compared
| Model | Provider | Price (input / output, USD per 1M tokens) | EU processing | Assessment for German |
|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic (USA) | 4 / 20 | via AWS Bedrock or Google Vertex in EU regions | regularly at the top of style rankings; strong on long, nuanced texts |
| Claude Sonnet 5.5 | Anthropic | 2 / 10 | same as Opus | very good value for everyday texts |
| Claude Haiku 5.5 | Anthropic | 0.10 / 0.50 | same as Opus | low cost for large volumes, somewhat simpler style |
| GPT-6.1 Sol | OpenAI (USA) | 2 / 10 | EU data residency for new API projects | solid German, very versatile |
| GPT-6.1 Terra | OpenAI | 1 / 5 | same as Sol | cheaper option for standard texts |
| Gemini 4 Argon | Google (USA) | 2 / 10, from Dec 1, 2026 4 / 20 | via Vertex AI in EU regions such as Frankfurt | #1 in Arena’s German ranking |
| Gemini 3.8 Flash | 0.75 / 3.75 (promotional until year-end) | same as Argon | fast and cheap for volume | |
| Mistral Large 4 | Mistral (France) | 0.70 / 2.10 (preview) | European data centers | European option; no published German ratings |
| DeepL | DeepL (Cologne) | per character | EU provider | translation specialist, with formal/informal control and glossaries |
| Open models (e.g. Mistral Small, Gemma) | various | own infrastructure | in your own data center | for maximum data protection; some surprisingly strong on grammar tests |
On German models: projects such as Teuken-7B (OpenGPT-X) or LLäMmlein (University of Würzburg) matter for research, but aren’t competitive for business texts today. In July 2026, Soofi S, a new German-English open model, was added; so far it’s only available in a closed beta. Heidelberg-based Aleph Alpha signed a merger with Canadian provider Cohere in September 2026; which models will result is still open.
Which business platforms provide these models in a GDPR-compliant way is compared in our article on ChatGPT alternatives for business.
How to recognize good business German
Whether an AI text is good can be checked against ten criteria. This list also works as the basis for your own test.
| Criterion | Typical AI mistake | What it should look like |
|---|---|---|
| Anglicisms | „Wir haben das Meeting gecancelt und warten auf Feedback.“ | „Wir haben die Besprechung abgesagt und warten auf Rückmeldung.“ |
| Sentence structure | English word order, many participles: „Basierend auf Ihrer Anfrage …“ | „Auf Ihre Anfrage hin …“ or „Sie haben gefragt, ob …“ |
| Nominal style | „Die Durchführung der Prüfung erfolgt durch uns.“ | „Wir prüfen das.“ |
| Sie/du | switching mid-text | one consistent form of address |
| Gender-inclusive language | mixed forms in the same text | one consistent variant per house style |
| Quotation marks | ”English marks” | „German marks“ |
| Numbers and dates | 1,250.50 €, 11/12/2026 | 1.250,50 €, 12.11.2026 |
| Stock phrases | „In der heutigen schnelllebigen Welt“, „Es ist wichtig zu beachten“ | straight to the point |
| Compound words | „Kunden Service Prozess“ | „Kundenserviceprozess“ or „Kundenservice-Prozess“ |
| Region | „ß“ in Swiss texts, „Januar“ in Austrian texts | „ss“ in Switzerland, „Jänner“ in Austria, depending on audience |
Recommendations by task
| Task | Recommendation | What to watch |
|---|---|---|
| Marketing copy, website, newsletter | Claude Opus or Sonnet, Gemini | stock phrases and anglicisms; specify tone |
| Customer emails and complaint replies | any top model with a house style prompt | consistent form of address, empathy without exaggeration |
| Product descriptions at scale | cheaper tiers (Haiku, Flash, Terra) | consistent terminology, number formats |
| Summaries of German documents | top model with a large context window | faithfulness to the original, no invented details |
| Technical and legal texts | top model with access to your sources (RAG), reviewed by experts | facts, citations, current legal situation |
| Translation English → German | DeepL or a top model with a glossary | terminology, Sie/du, translationese |
| Easy-to-read German (Leichte Sprache) | specialized fine-tuned models or a top model with the rulebook | rules per DIN SPEC 33429, testing with the target group |
| Confidential content | Mistral, open models in your own data center | plan for some quality trade-off and more editing |
A 2025 study on translation into easy-to-read German found that specifically fine-tuned smaller models outperformed a large general-purpose model on public-sector texts. For specialized tasks, the biggest model isn’t automatically the best.
The biggest lever: a system prompt for your German house style
The differences between top models are small. The difference between a text without guidance and a text with a clear house style is large. Anthropic explicitly recommends in its documentation to set the target language in the system prompt and to ask for idiomatic text “like a native speaker”.
A good system prompt for German texts defines:
- form of address (Sie or du) and gender-inclusive language
- region (Germany, Austria, Switzerland)
- tone and text type
- how to handle anglicisms and typical AI phrases
- German typography for quotation marks, numbers and dates
- how to handle missing facts
Build your house style and copy the finished prompt:
Free prompt generator
Your German house style as a system prompt
Choose your settings, copy the prompt and add it to ChatGPT, Claude, Gemini or your company LLM.
Your system prompt
Tip: Add two or three short sample texts from your own communication and a list of your technical terms. Examples work better than any rule.
Consistent German across all your company's AI texts
We set up your house style in your company LLM, choose the right model for your texts and train your team to write with AI.
Data protection: which models can you use?
For business texts containing customer data or internal information, GDPR compliance depends less on the model than on how you access it. What matters is a data processing agreement, excluding training on your data, and where processing happens.
- OpenAI: EU data residency for new API projects; Business and Enterprise plans with a data processing agreement.
- Google Gemini: processing via Vertex AI in EU regions such as Frankfurt is possible. Don’t use the global endpoint.
- Anthropic Claude: we found no EU option in Anthropic’s own API. Via AWS Bedrock or Google Vertex AI, Claude is available in European regions.
- Mistral: runs in European data centers. A data processing agreement is still required; compliance isn’t automatic.
- Open models: in your own data center, data never leaves the company. In return, you’re responsible for operations and quality assurance. When that pays off is covered in our local AI vs. cloud AI comparison.
What does a German AI text cost?
List prices per million tokens feel abstract. Converted to a typical text: a 150-word product description in German is roughly 250 to 300 output tokens, plus prompt and house style as input. Even with a top model, the pure model cost is a fraction of a cent to a few cents per text.
Two points matter more than the list price:
- The German surcharge: depending on the tokenizer, the same content needs 15 to 60% more tokens in German than in English. At large volumes, that adds up.
- Editing time: the real cost arises when employees have to correct texts. A cheap model whose texts need reworking every time is more expensive than a top model whose texts are almost ready.
Technical and legal texts: beware of invented facts
Good German doesn’t mean correct content. Especially in specialist texts, language models occasionally invent details that sound convincing.
- In July 2025, the Cologne Local Court criticized a lawyer’s brief that contained invented or incorrect citations (case no. 312 F 130/25).
- A 2026 study on questions about German statutory law shows that models without access to current sources perform significantly worse on legal changes after their training cutoff. Matching against versioned statutes noticeably improved the results.
For specialist texts: give the model the sources, for example via a knowledge base, and have experts check the facts. Since October 2026, Google has also explicitly required that AI-generated content be manually fact-checked before publication.
Labeling: The EU AI Act doesn’t require blanket labeling of AI texts. There is an obligation for texts that inform the public on matters of public interest, unless a human has reviewed them and holds editorial responsibility. More in our EU AI Act compliance checklist.
How to test which model suits you
Since there’s no reliable style benchmark, a small test of your own is the best basis for a decision. It takes a few hours:
- Choose tasks: five typical texts from your daily work, such as a customer email, a product description, a newsletter paragraph, a summary and a specialist text.
- Same conditions: use the same house style prompt and the same tasks for all models.
- Rate blind: give the texts without model names and in random order to two colleagues who write well in German.
- Check against criteria: the ten criteria from the table above, plus the question “Would you send this text as is?”
- Measure editing time: how many minutes until the text is ready to send? That’s the most important metric.
Avoid having an AI do the rating. AI judges demonstrably prefer texts from their own model family and tend to overrate non-English texts.
Checklist: checking the German quality of AI texts
Checking the German quality of AI texts
Ten criteria for business texts plus ground rules for company use
Language
- No unnecessary anglicisms Besprechung instead of Meeting, Rückmeldung instead of Feedback
- German sentence structure, no translated sound
- Active and verbal rather than nominal style
- No typical AI stock phrases
- Compound words spelled correctly
Form
- One consistent form of address (Sie or du)
- Consistent gender-inclusive language per house style
- German quotation marks „…“
- Numbers and dates in German format 1.250,50 €, 12.11.2026
- Region considered (Germany, Austria, Switzerland)
Company use
- House style set up centrally as a system prompt
- Only accounts with a data processing agreement for company data
- Facts, figures and quotes checked before publication
- Own short test with five typical texts carried out
Frequently asked questions
There's no clear winner, because no benchmark reliably measures the style quality of German texts. In Arena's German ranking in November 2026, Gemini 4 Argon led, closely followed by a group including Claude Opus 5.5 and GPT-5.6. In a German business benchmark, Claude came first. The top models from Anthropic, Google and OpenAI all write good German; a clear house style prompt makes a bigger difference than the model choice.
Language models learn mostly from English text. In Common Crawl, around 5.8% of pages are German and 42.6% English. Many models also process tasks internally in English. This shows in anglicisms, English sentence structure, participle constructions and stock phrases. Concrete linguistic instructions in the prompt significantly reduce the effect.
All three are close for German texts. Claude and Gemini usually score slightly better in style and user-preference rankings, while ChatGPT is very versatile. For companies, data protection, existing infrastructure such as Microsoft 365 or Google Workspace, and costs often decide. Test with your own texts and a consistent house style prompt.
German models such as Teuken-7B or LLäMmlein matter for research but aren't competitive for business texts. In July 2026, Soofi S, a new German-English open model, was added; it's so far only available in a closed beta. For European data residency, Mistral from France is currently the most practical option.
Set the form of address explicitly in the system prompt, for example: address readers consistently with Sie and never switch to du. Store the prompt centrally in your company LLM or as custom instructions. For translations, DeepL has a dedicated formality parameter. The prompt generator in this article creates the right instruction.
Yes, if you specify the variant. Without guidance, models often mix forms. Decide whether you use neutral forms, paired forms, the colon, the asterisk or the generic masculine, and require consistent use throughout the text.
Yes. Language models bill by tokens, and German text, especially compound words, is split into more tokens. A study found around 58% more tokens in German with the GPT-4 tokenizer. Newer tokenizers narrow the gap; depending on the model, it's roughly 15 to 60%. Per individual text, model costs usually remain in the cent range.
Compliance depends on how you access the model, not on the model itself. You need a data processing agreement, an exclusion of training on your data and a suitable processing location. OpenAI offers EU data residency, Gemini runs via Vertex AI in EU regions, Claude via AWS Bedrock or Google Vertex in Europe, and Mistral in European data centers. Free chat accounts aren't suitable for company data.
For translation into German, DeepL is an excellent choice, with control over formal and informal address and glossaries for terminology. For writing new texts, summaries or replies, general language models such as Claude, GPT or Gemini are more versatile. Many companies use both.
Not across the board. The EU AI Act requires labeling for texts that inform the public on matters of public interest, unless a human has reviewed the text and holds editorial responsibility. There's no general obligation for marketing copy or customer emails. Always check facts, though, also because Google has explicitly required it since October 2026.
Conclusion: the model matters less than the house style
In 2026, the question of the best LLM for German texts has an uncomfortable answer: the top models from Anthropic, Google and OpenAI are so close that no leaderboard crowns a clear winner. Differences arise mainly from the task, data protection requirements and costs, and even more from how you instruct the model.
If you set up a clear house style as a system prompt, check facts and test with your own texts, you’ll get good business German from any top model. Without guidance, every model produces the familiar “translated” sound.
Good German from AI, in your house style
We choose the right model with you, set up your house style centrally and train your team so AI texts sound like your company.
