How good is AI in languages other than English?
Three probes that show whether a model holds up in your language, and what the model makers themselves put in writing about languages beyond English.
Élise runs a small design studio in Brussels. She works in French all day and has just taken on a supplier in Aachen who answers in German and nothing else. She needs to answer two questions. Is this thing any good in French, a language she can hear well enough to grade? And can she rely on it in German, which she cannot?
Those two need different methods. Most writing on the subject answers a third question, which language is best, and that one has no honest answer: the comparison has never been published for the 59 models we carry, one language at a time, and a ranking assembled without it would be invention. So here is Élise's afternoon instead.
Start with the language you can grade
All three parts stay in French.
First she asks for something idiomatic, a real request whose answer cannot be assembled out of a dictionary. Write me two sentences cancelling an appointment at the last minute without sounding like someone who does not care. A model with a grip on French produces the small softeners a French speaker reaches for without thinking. A model working from English templates produces French-shaped English, polite in a way nobody in Brussels is polite.
Second she asks it to explain two words out of her own French, "la drache" for a downpour and "septante" for seventy, and where each of them is used. This probe separates knowing a language from knowing the variety of it you speak. A strong answer places both in Belgium and can say what a reader in Paris would reach for instead. A weak answer either invents a meaning with complete composure, the behaviour our piece on AI hallucinations describes, or treats her own French as a mistake to be corrected.
Third she asks for a short formal note. Six lines, a named reader, a stated purpose: a reminder to a client whose invoice is three weeks late. The brevity is the point, because six lines gives a model nowhere to hide behind fluency, and formal register is where the seams show first.
Read all three replies before deciding anything, then put the same requests to another model. Changing model part way through a conversation needs no retyping, so the exercise fits in one thread.
The tells of something translated out of English
What Élise is reading for has a signature, and once you have seen it you cannot unsee it. Sentences come back grammatical and slightly too long. Connectives get spelled out where a French speaker would let two clauses sit next to each other. Set phrases arrive as literal renderings, so an email opens by hoping the message finds you well, which is not how anything in her own sent folder has ever opened. Register drifts upward on its own, turning a note to a colleague of five years into correspondence from a notary. Idiom goes flat, swapped for the safest available synonym.
None of that is an error a spellchecker can see, which is why proofreading tools miss it and a native reader clocks it in the first line.
German was the harder half
In German Élise cannot mark the output, so the method changes. Everything has to come from evidence she already holds or a second opinion she can read. She writes what she means in French, has it put into German, and sends that German to a second model from a different lab for a literal French rendering. Where the two French versions part company is where she has something to ask about, and that is all the detour buys her: it cannot tell her the German is good, only where to look.
She can do that by switching model inside the same thread, or by branching the message into its own conversation to keep both versions. A branch opens as a new thread instead of a second panel, so comparing means moving between two conversations.
Then the probe that did the most work in French does it again in German. She asks the model to account for its own choices: why "Sie" rather than "du" for a supplier she has never met, why that particular word for an invoice, which of her French phrases had no German equivalent. She cannot judge the German. She can judge the reasoning about the German, and she can check three named words against the supplier's own emails, already sitting in her inbox.
Which makers put a language claim in writing
A few makers publish their language claims, and those are more use than a ranking. The claims below were checked against the makers' pages on 2 October 2026.
Mistral, a French lab, names the languages on its model card for Mistral Large 3: dozens supported, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean and Arabic. Alibaba's Qwen cards are the most specific of the four about breadth. The card for Qwen3 235B A22B claims "Support of 100+ languages and dialects" along with multilingual instruction following and translation, and the Qwen3-VL card puts a figure on reading text inside images: 32 languages, up from 19. Google states that its own "Gemini web app is currently available in more than 70 languages and over 230 countries and territories". OpenAI publishes a list of 59 languages that the ChatGPT interface can be set to.
Those last two belong in a different category from the first two.
A published language list is a support claim rather than a quality measurement. Mistral's and Alibaba's lists describe what the model was built to handle. Google's and OpenAI's describe what their own apps are available in, which is a claim about a product and says nothing about a model. None of the four tells you whether the German in your letter will read as though a German wrote it. They tell you which languages a maker will stand behind in public, which is a good place to start looking and a bad place to stop.
So the claims narrowed where Élise looked. They settled nothing. The three probes settled it, in about the time it took her to make coffee.
The interface is English only
One thing to be plain about, because it is the first thing people notice. Our own menus, buttons and settings are in English, and only English. There was no language picker in the app at our October check, and nothing reads your system language to adjust the interface. Next to what OpenAI documents for its own interface that is a real gap, and we would rather write it down than let you find it.
It has no bearing on what the models do. Write in French and the answer comes back in French. Write in German and it comes back in German. The English stops at the furniture.
One mechanical detail, since it surfaces whenever documents get long. Models count text in tokens, and how many tokens a word costs is not constant across languages. OpenAI's own help page on tokens gives a rough figure for English, then says outright that "other languages can have different relationships between characters, words, and tokens". It does not say which direction, and neither will we, having not measured it across 59 models. A long document in your language may fill a context window at a different rate than its English equivalent, so judge length by the answers you get and not by the page count.
Writing the letter
The reminder went to Aachen late that afternoon, and most of what made it usable had nothing to do with which model produced it.
Élise named the reader and the relationship inside the request, because left alone the output drifts formal, and a cold letter to a supplier she wants to keep for five years is a small own goal. She handed it the German conventions she already had, the greeting and the sign-off lifted from the supplier's own messages, and told it to use those rather than invent equivalents. And she deleted the opening line it volunteered, the one hoping the message found the recipient well, because nothing in the supplier's own letters had opened that way.
One more thing, before you paste correspondence that belongs to somebody else. We do not train on your conversations. What the provider behind a given model does with them varies by model, and the models where we know of a difference carry a notice on their info card in the picker.
Where we would land
So, how good is AI in languages other than English? Good enough that Élise sent the letter. Not good enough that she sent it unread.
That is the judgement, and it is the only shape available, because the language-by-language gradient that would settle it has never been published. Where a language does something English does not, in its politeness system or its regional varieties, read that as a place to slow down. A difference like that forecasts nothing about the quality you will get. What the model cannot do is tell you which of those situations you are standing in. That part comes from you, or from someone you trust with the language.
The three probes above are more useful for your work than a published language list. Twenty-two of the 59 models we carry are open on a free account, and the full set runs across 17 providers, so putting the three probes to four of them teaches you more about your own language than a ranking can. Do that before you decide anything, including whether you want to pay us.