Every major AI chat service understands Dutch. That stopped being a differentiator a while ago, and it's not where people get stuck. Things go wrong in the other direction: in the text that comes back. It's usually grammatically flawless and yet recognizably un-Dutch, because the sentence structure, the expressions and the tone are borrowed from English.
That difference is surprisingly easy to pinpoint. It's in anglicisms nobody says out loud, in expressions translated word for word, in "je" and "u" (informal and formal "you") getting mixed up within a single answer, and in an enthusiastic tone that seems to come straight from an American newsletter. We tested nine services on exactly those points and gave each provider a score from 1 to 5.
The short outcome: Claude is the only one to score a 5. Four services are on a 4 and are perfectly usable as long as you do an editing pass yourself. Four get stuck on a 3.
How we tested the Dutch
All nine services got the same four assignments, written in Dutch and without extra instructions about style. That last point matters. Almost every model writes better Dutch if you explicitly ask for it, but most people don't ask. We wanted to know what you get when you simply type in your question.
The four assignments:
- A business email to a customer who pays late. This tests the form of address, the polite formulas and whether the model stays in the Dutch register.
- A 400-word blog post on an everyday topic. This tests rhythm, heading style and the tendency to put everything in threes.
- A translation of an English paragraph containing two idioms. This tests whether the model translates literally or looks for the equivalent.
- An informal conversation with colloquial language and a joke in it. This tests a feel for everyday spoken Dutch.
We assessed idiom, consistency in the form of address, tone and how much editing is needed before you'd send the text. We didn't test comprehension: all nine understand a Dutch question without difficulty, even with typos or a dialect word in it.
The scale: a 5 means you can send the text as is, a 4 that you still need to fix something in each paragraph, and a 3 that you get usable raw material instead of a finished text. No scores below 3 occurred.
The scores of nine services
| Provider | Score | Verdict |
|---|---|---|
| Claude | 5 | Most natural Dutch: correct idiom, no anglicisms, steerable tone |
| ChatGPT | 4 | Grammatically fine, but an American tone and the occasional literal translation |
| Gemini | 4 | Correct and strong at translation, but flat and businesslike |
| Perplexity | 4 | Answers well in Dutch and finds Dutch sources; translated quotes are awkward |
| Le Chat | 4 | Surprisingly good, neutral tone, stumbles over colloquial language |
| Copilot | 3 | Understands Dutch but falls back into English in long answers |
| Grok | 3 | Fine Dutch, but the humor doesn't land |
| DeepSeek | 3 | Understands Dutch, produces it less fluently |
| Poe | 3 | The models speak Dutch, the app, support and invoices don't |
Two things stand out. The field is close together, and the standout at the top has only one name. In addition, Le Chat from the French company Mistral scores as high as ChatGPT and Gemini: a European provider isn't automatically better at Dutch, but it isn't worse either, which is a persistent misconception among people who are looking at European services for other reasons.
What exactly goes wrong in Dutch AI texts
The mistakes are strikingly consistent. Four categories explain almost everything you have to take out.
Anglicisms nobody says out loud
It's rarely about actual word errors. It's about constructions that technically exist but that no Dutch speaker uses: "we zijn opgewonden om te delen" (a literal "we're excited to share"), "voel je vrij om contact op te nemen" ("feel free to reach out"), "dit zal je helpen leveren op je doelen" ("this will help you deliver on your goals"), "laat me weten als je vragen hebt" ("let me know if you have any questions"). Also popular are verbs borrowed straight from English and given a Dutch ending, such as geïmpacteerd ("impacted"), gealigneerd ("aligned") and geëscaleerd ("escalated") in the sense of opgeschaald ("scaled up").
Expressions translated word for word
This is the clearest sign that a text comes from a model. "At the end of the day" becomes "aan het eind van de dag" (which in Dutch only means literally at the end of the day) where you mean "uiteindelijk" ("ultimately"). "It's not rocket science" becomes "het is geen raketwetenschap". "Touch base" becomes something involving a base. In the translation assignment, this went wrong with five of the nine services; Claude and Gemini did choose the Dutch equivalent.
"Je" and "u" in the same answer
The form of address is the weakest spot of almost every model. English doesn't have the distinction, so the choice gets made again for every sentence. A typical answer opens with "Geachte heer De Vries" ("Dear Mr. De Vries") and the formal "u", then switches to the informal "je" three paragraphs later in a bulleted list. The reverse happens too: an informal chat that suddenly says "kunt u dit bevestigen" ("could you confirm this", in the formal form) halfway through. This is the mistake readers notice fastest, and the one you can most easily prevent by fixing the form of address in your prompt.
The American tone
Dutch business language is more down-to-earth than American. A model that opens with "Geweldige vraag" ("Great question"), ends every list with a word of encouragement and closes the text with "ik hoop dat dit helpt" ("I hope this helps") doesn't sound friendly in Dutch but fake. Then there's the notation: dates like 08/10/2026 instead of 10 augustus 2026 (10 August 2026), amounts with a period as the decimal separator instead of the Dutch comma, and headings in which every word is capitalized.
The verdict per provider
Claude is the only service that delivers texts you can send without intervention. The idiom is right, the form of address stays consistent within an answer and the tone can genuinely be steered: ask for dry and businesslike and you get dry and businesslike, not the same answer with different adjectives. What else sets Claude apart is in our explanation of Claude.
ChatGPT writes grammatically flawless Dutch. It loses its point on tone: it's the service that most often sounds enthusiastic where that doesn't fit, and that occasionally translates an expression literally. With a style instruction added, it comes close to Claude.
Gemini is correct and reliable, and the best of the bunch at translation. The drawback is the flat, businesslike tone: texts are factually in order but have little rhythm of their own, so they quickly start to resemble each other.
Perplexity answers well in Dutch and also finds Dutch-language sources, which makes it pleasant for research. Quotes it translates from English sources are awkward.
Le Chat from Mistral surprises with a neutral, tidy tone and few anglicisms. Where it stumbles is colloquial language: informal assignments produce correct but stiff Dutch.
Claude
Anthropic · United States
€21.02 /mo
$20 list price
The writer and programmer of the bunch. Long documents, natural language and Claude Code included with Pro.
- Free
- Claude Sonnet 5 as the default model, with a daily message limit.
- Dutch
- 5/5
- Data
- United States and EU regions via cloud partners
Le Chat
Mistral AI · European Union
€15.76 /mo
$14.99 list price
The European answer. Fast, cheap and with processing inside the EU.
- Free
- Generous free use with web search and image generation.
- Dutch
- 4/5
- Data
- European Union (France and Sweden)
Copilot understands Dutch fine, but falls back into English in longer answers, sometimes in the middle of a list. Grok writes correct Dutch, but the humor it's known for doesn't come through. DeepSeek understands the question but produces less fluent Dutch. With Poe, the models speak Dutch just fine, but the app, the support and the invoices are in English.
Which model do you actually get when you type in Dutch
The name of the service doesn't tell the whole story, because behind every chat is a model that differs per subscription. That matters, because sometimes you already get good Dutch without paying.
With Claude, the free version runs on Claude Sonnet 5, the same model that does the daily work in the paid plan. That's the most direct route to good Dutch for zero euros. With ChatGPT, GPT-5.6 Luna has been the default model for free users and for the Go subscription since 6 August 2026; the heavier GPT-5.6 Sol is in Plus. With Le Chat, you work with Mistral Medium 3.5.
| Model | From | Released | Context | Dutch | Best for |
|---|---|---|---|---|---|
| Claude Sonnet 5 | Anthropic | 30 June 2026 | 1M tokens | Excellent | Daily writing, coding and agents |
| GPT-5.6 Sol | OpenAI | 9 July 2026 | 1,05M tokens | Good | Heavy analysis, agents and professional work |
| GPT-5.6 Luna | OpenAI | 9 July 2026 | 1,05M tokens | Fair | Quick questions and high volumes |
| Gemini 3.5 Flash | 19 May 2026 | 1,05M tokens | Good | Fast multimodal work and large volumes | |
| Mistral Medium 3.5open | Mistral AI | April 2026 | 256K tokens | Good | European deployments and agent work |
If you want to know which model does which task best, the task-based comparison of GPT-5.6, Claude 5 and Gemini 3.5 is the logical next step. If you're purely looking for the free route, take a look at free AI chat in Dutch.
How to get better Dutch out of any model
The difference between a 3 and a 4 is often in your prompt. Four interventions work with all nine services.
Write your assignment in Dutch. Models adopt the language of your prompt as a style example, so an English instruction with "answer in Dutch" produces translated-sounding Dutch more often than a prompt written entirely in Dutch.
Fix the form of address. One sentence is enough: write in the "je" form and never use the "u" form. This is the cheapest quality gain there is.
Include a sample text. Paste a paragraph you wrote yourself earlier into the prompt and ask for that same tone. That works better than ten adjectives about style.
Have the model check its own text. In a second message, ask it to remove all anglicisms and literally translated expressions. Models are considerably better at recognizing those mistakes than at avoiding them.
Anyone who seriously produces Dutch texts rarely gets by with one model. You write in Claude, you check facts in Perplexity and you have Gemini summarize a PDF. That normally means three subscriptions, and you can see how that bill adds up in our comparison of all AI subscriptions. In that case, one subscription that combines those models saves the most.
Frequently asked questions
Which AI writes the best Dutch?
Claude writes the most natural Dutch of all the major chat services. The idiom is right, hardly any anglicisms creep in and the tone is easy to steer. ChatGPT, Gemini, Perplexity and Le Chat follow closely, but their texts usually still need an editing pass.
Why does AI Dutch sound so English?
Most mistakes aren't grammar mistakes but style mistakes that come from English: anglicisms based on 'reach out' and 'excited', expressions translated word for word, and an enthusiastic tone that comes across as exaggerated in Dutch. The notation is often wrong too, with American dates and periods instead of commas in amounts.
Can I use a free AI that writes good Dutch?
Yes. The free version of Claude runs on Claude Sonnet 5, the same model that does the daily work in the paid plan, and it's also the best Dutch writer of the bunch. The free versions of ChatGPT, Gemini and Le Chat also deliver usable Dutch, with limits on uploads and images.
How do I get better Dutch texts out of an AI?
Write your prompt in Dutch, explicitly choose your form of address, include a sample text of your own and ban anglicisms with a short list of words you don't want to see. Then have the model check its own text for literally translated expressions. That lifts almost any model by half a point.