Best AI for German: GPT tops a 17 model test from €22.99 a month, and 15 of 17 models missed one proverb

Which AI is best for German?

Gemini Pro tops the biggest public German leaderboard, a multiple choice test, while GPT wins when the model has to write the German answer itself. Here are the German scores from our Multilingual LLM Bench, run on 2 October 2026, with what one full run of 150 questions in 30 languages cost each model:

Of the 17 models we ran, only GPT-6 Sol and GPT-6 Astra got all five of them right. Eight models got four, and Claude Opus 5.5 and Gemini 3.1 Pro are two of those. They're one answer behind. Five questions is a small sample, so one answer is enough to move a score by 20 points. That's why it helps to look at a bigger test too. Artificial Analysis tests German with Global-MMLU-Lite, which is a multiple choice test. It puts Gemini 3.1 Pro first in German at 95, with four Claude Opus models tied at 93, as we read it on 18 October 2026. It doesn't have a GPT-6 or a Claude Opus 5 model on the board yet.

How did 17 models do on Whizi's German questions?

GPT-6 Sol and GPT-6 Astra were the only two models that got all of the German questions right. Each model got the same five questions in each of 30 languages, at low reasoning, on 2 October 2026. We graded every answer against a key with a source behind it, and an answer in the wrong language scored zero. All of this is from the October 2026 run of our Multilingual LLM Bench.

Treat it as a rough order. One question is a fifth of the score.

If you split the answers by question, most of the misses pile up on one of them:

The public question is the one you can see. It asks, in German, for the proverb that says that if you talk big, you have to deliver. The answer, as Duden's idiom dictionary and German Wiktionary both give it, is "Wer gackert, muss auch ein Ei legen": whoever cackles has to lay an egg. Only GPT-6 Sol and GPT-6 Astra wrote it.

Five of the models answered "Wer A sagt, muss auch B sagen" instead, and all three of the Claude models were among them. That one means something else: once you start, you have to finish. Gemini 3.1 Pro came the closest. It wrote "muss auch Eier legen", which is one word off, and our strict key marked it wrong. GLM-5.3 used up its whole budget of 8,000 tokens on thinking and returned two asterisks.

Both of the regional questions are about Austria and Switzerland, not Germany. The models got them right 15 and 12 times out of 17. So check any Swiss or Austrian fact twice, even when the German in the answer reads perfectly.

German was the easiest language in the whole run. The 17 models averaged 64.7% in German, which is ahead of English at 61.2% and far ahead of Spanish at 25.9%. That tells you more about how hard each set of five questions was than it does about the language. So don't use our bench to rank German against Spanish.

Is ChatGPT good at German?

Yes, ChatGPT is good at German. OpenAI's simple-evals results put its o3 reasoning model at 0.905 in German, on an MMLU test that professional translators rewrote in 14 languages. GPT-4.1 mini, which is a smaller model, scored 0.823. The GPT that you get depends on your plan. OpenAI's help page on GPT-6, which we read in October 2026, says that Free and Go default to GPT-5.6 Luna. Plus gets GPT-6 Sol in Chat, and you only get Astra in Chat on Pro.

Claude and Gemini are close behind it, or level with it. Anthropic's multilingual support page says that Claude Sonnet 4.5 kept 97.0% of its English score in German, and that the smaller Claude Haiku 4.5 kept 94.3%. On Artificial Analysis, Gemini 3.1 Pro scores 95 in English and 95 in German, so it doesn't lose a thing in German on that test. Claude Opus 4.6 drops two points, from 95 to 93.

Google's plan page now lists Gemini 4 Argon on Google AI Pro. It's newer than our run, so we haven't tested its German.

Is DeepL better than ChatGPT for German translation?

DeepL is the better pick for a straight translation, and ChatGPT or Claude is better when the translation also has to be rewritten. DeepL is a company from Cologne, and it runs a language model of its own that it built for translation. In January 2025 it said that blind tests with linguists found a 1.4x improvement for English and German over its previous model. It also says that in its 2024 blind tests, ChatGPT-4 needed 3x as many edits as DeepL did. Those are DeepL's own tests, and they compare it with GPT-4, not GPT-6.

DeepL Pro costs €7.49 a month for one person, billed by the year with VAT included, as of October 2026. That buys you 300,000 characters a month. The free plan gives you 50,000.

When you use a chatbot to translate, give it the rules that DeepL can't take. This prompt sets the form of address, the German number format and the gender style:

```

Übersetze diesen Text ins Deutsche, für [Kunden / Kollegen / eine Behörde].

Sprich die Leser durchgehend mit [Sie / du] an.

Behalte alle Namen, Zahlen, Daten und Beträge genau bei.

Schreib Zahlen deutsch: 1.234,56 € und 18.10.2026.

[Kein Gendern / Gendern mit Doppelpunkt, zum Beispiel Kolleg:innen]

Wenn eine Redewendung kein deutsches Gegenstück hat, übersetze die Bedeutung.

Nenne am Ende jedes Wort, bei dem du unsicher warst.

[Text hier einfügen]

```

Our guide to using AI to translate has the checks to run after.

How do you get an AI to use Sie or du?

Set it once in the settings, then you don't have to ask in every chat. DeepL is the only tool here with a switch for it. Its formality feature offers Formal and Informal for German on DeepL Pro, so "Vielen Dank, dass Sie DeepL Pro verwenden!" becomes "Danke, dass du DeepL Pro benutzt!".

Chatbots don't have a switch, but each has a place for a standing rule. ChatGPT takes it in custom instructions, on every plan, up to 1,500 characters on Free and Go and 5,000 on Plus. Claude takes it in profile preferences. Gemini takes it in Personal context. Paste this in:

```

Antworte immer auf Deutsch und sieze mich.

In Texten für Kunden: Sie. In Texten fürs Team: du.

Schreib deutsche Anführungszeichen („…“) und deutsche Zahlenformate.

Für Leser in der Schweiz: ss statt ß.

```

The last line matters for Swiss readers. Swiss German spelling drops ß, and the models default to the German standard.

How much do ChatGPT, Claude and Gemini cost in Germany?

ChatGPT Plus costs €22.99 a month in Germany's App Store, Claude Pro €22.00 and Google AI Pro €21.99, read on 18 October 2026. Google AI Pro is a cent cheaper than Claude Pro and a euro cheaper than ChatGPT Plus.

The cheapest plan with a model that got all 5 of our German questions right is ChatGPT Plus, at €22.99. The free tiers are a step down from that. OpenAI gives its Free users GPT-5.6 Luna, Anthropic gives Claude Free users Sonnet and Haiku, and Google gives its free Gemini users the Flash-Lite model. For every country, see ChatGPT Plus prices by country and Claude Pro prices by country.

Which AI keeps your data in the EU?

None of the big three consumer plans does. As of October 2026, OpenAI, Anthropic and Google sell an EU data region only to business, school or cloud customers. Here's where each one stands, compiled from each vendor's own docs on 18 October 2026:

Mistral Vibe is the European chatbot that most people have in mind, and it was called Le Chat until Mistral renamed it in May 2026. Mistral's privacy policy says: "We prioritize selecting providers within the European Union that strictly adhere to the GDPR." That's a preference. It isn't a promise of where your chats will be kept. Mistral's pricing page lists Vibe Pro at $14.99 a month before tax.

Aleph Alpha, the Heidelberg lab, no longer sells a chatbot to the public. On 16 September 2026 it signed an agreement with Cohere to form one transatlantic company, and the deal still needs regulatory approval. Until it closes, Aleph Alpha says it runs independently. It released Kolibri on 3 October 2026: 78 billion parameters, about 3 billion active per token, with German at about 23% of its training data. The weights are free under Apache 2.0. Aleph Alpha's own German score for it is 70.8, a figure nobody else has checked yet.

Which AI is best for learning German?

For speaking practice, Duolingo Max has an AI video call in German, and any big chatbot handles grammar and writing. Duolingo's blog said on 14 January 2026 that Video Call is open to Max subscribers learning German on iOS. A chatbot can't grade your accent, but it explains a grammar rule well. Ask it to correct your text, mark each mistake and name the rule it breaks: Dativ, Akkusativ or word order. The learning apps that rank for this search sell the same idea in a tutor shape. Langua offers German among 16 languages, with AI conversation that corrects each mistake and explains it. Talkpal teaches German too, and its Basic plan is free.

Where does Whizi fit for German?

Whizi is our product, so here's where it helps and where it loses. Pro at $29.99 a month includes ChatGPT Sol, Claude Opus, Gemini Pro and Mistral Large, and its side by side model chats run one German prompt through two of them at once, each billed at its own rate. An always on Skill can hold your Sie or du rule, in up to 4,000 characters, so you set it once. GPT Astra needs Powerhouse at $49.99, and Starter at $15.99 has ChatGPT Luna and Claude Haiku. Whizi loses on privacy if you use it for German business work. It offers no EU data region, and its privacy policy names Canada and the United States as the places where processing may happen. It doesn't have DeepL either. Side by side model chats shows how compare works.