ChatGPT alternative for Dutch: which AI speaks our language?

Every major model can write Dutch. Since 2025 that is no longer in question. The question that matters now is more subtle: does the model write Dutch, or does it translate English into Dutch? You won't see that difference in a language test, but you will see it in the fifteen minutes you spend rewriting an email that feels "translated".

We tested the eight best-known providers on Dutch and gave them a score from 1 to 5. Claude is the only one to come out with a 5. ChatGPT gets a 4: solid, but with an American undertone you have to edit out. At the bottom are Copilot, Grok and DeepSeek with a 3: they understand Dutch, but produce it less smoothly.

Below are the full scores and, more importantly, which mistakes you'll run into with which model, so you know what to watch for.

How we tested

Our scores come from our own hands-on work, not from a standardized benchmark. That is the most important caveat right away: it is a judgment, not a measurement with two decimal places. We used the same tasks for every provider: a business email, a persuasive paragraph, a summary of a Dutch document, a translation from English into Dutch, and a conversation in colloquial language.

What we looked at: does the model choose the right register, does it use Dutch expressions instead of translated English ones, does it stick to the form of address, and can you adjust the tone when you ask. A more extensive version of this test with more examples is in which AI speaks the best Dutch.

The scores per provider

Note that last row, because it is about something else. With Poe the model isn't the problem: you can talk to the same top models there. It is the environment that is in English: the interface, customer service and invoices.

Where Dutch AI texts go wrong

The mistakes we came across fall into three categories. None of them is a grammar mistake, and that is exactly why they slip past a spell checker.

Anglicisms that look Dutch

This is the most common and the most treacherous mistake, because there is no language error in it. The model writes a correct Dutch sentence that nobody would actually say.

Typical examples from our tests: "Ik hoop dat dit helpt" (I hope this helps) as the closing line of an email, "Aarzel niet om contact op te nemen" (don't hesitate to get in touch), "Laten we het opsplitsen in drie stappen" (let's break it down into three steps), "Dat is een geweldige vraag" (that's a great question). All direct transfers of standard English phrases. Word choice too: "impacteren" (to impact) instead of "beïnvloeden" (to influence), and "delivered" becoming "geleverd" (supplied) where "afgerond" (completed) is meant.

ChatGPT does this most often of the strong models, and it is the main reason it stays at a 4 instead of a 5. Claude avoids it almost entirely: it chooses closings like "Laat het even weten" (let me know) or "Ik hoor het graag" (I look forward to hearing from you), which is what a Dutch letter writer would do too.

Je or u, and the switch halfway through

English "you" makes no distinction between formal and informal. A model translating from an English thought structure therefore has to guess, and when in doubt it often guesses the formal "u", because that looks safer. That makes texts for a consumer audience needlessly distant.

The switching is worse. Ask for a long text using the informal "je" and many models start out neatly, but slide back to "u" in the second half, often exactly when the subject becomes a bit more formal. Claude and Le Chat hold the form of address best. Gemini is consistent, but consistently businesslike.

Literally translated expressions

This is where it goes visibly wrong. English figures of speech that land in Dutch word for word: "aan het einde van de dag" (at the end of the day) where "uiteindelijk" (ultimately) is meant, "de bal ligt bij jou" (the ball is in your court), "we zijn op dezelfde pagina" (we're on the same page), "laaghangend fruit" (low-hanging fruit). Some of these have become business jargon by now, but in a customer email they stand out like a borrowed coat.

Conjugations of verbs that aren't Dutch verbs, such as "geüpdatet" (updated), "gedownsized" (downsized) and "gebriefd" (briefed), also show up structurally with the weaker models. Copilot also falls back on English in long answers: it starts in Dutch and ends in English, usually in the list at the end.

Why Claude writes better Dutch

We can't look into Anthropic's training data, so this remains an observation rather than an explanation. What we do see: Claude treats Dutch as a target language rather than as a translation. It chooses Dutch sentence structure with the subordinate clause in the Dutch position, uses Dutch conjunctions, and holds register and form of address across long texts.

Also relevant in practice is that Claude responds well to tone instructions. Ask for "businesslike but not stiff", "informal without becoming chatty" or "shorter and more direct", and the difference is there immediately. With Gemini the tone stays flat, however you ask.

That makes Claude Pro the strongest choice for anyone who writes a lot of Dutch: $20, so €21.02 per month including VAT, or €17.52 per month with annual billing. Claude Code and Claude Cowork are included and training is off by default on paid consumer plans. More background on the company and the model is in Claude AI explained.

What it costs to get good Dutch

The most striking row is the first. The best Dutch on the market is in a free account, because Claude Sonnet 5 also runs on the free tier. You give up limits and Claude Code, not language quality. If you don't want to create an account at all, Le Chat is an option; see free AI chat in Dutch for the full list of free Dutch-language options.

When you need more than one model for Dutch

In practice, people who write a lot choose per task. Claude for the text that has to go out, Perplexity when the facts need to be right and Dutch sources belong with them, Gemini for summarizing a two-hundred-page Dutch report, ChatGPT when images are needed.

That is sensible behavior with an expensive outcome: ChatGPT Plus and Claude Pro side by side is already €42.04 per month, and with Google AI Pro added €64.03 per month or €768.36 per year. That is why we bundle the models in one Dutch subscription of €12.95 per month (€129 per year), with a free tier without a credit card and invoices in euros including VAT. A side-by-side comparison of all subscriptions is on the comparison page.

If besides language you also want to know where your data is processed, that is a second consideration that often points the same way: see European ChatGPT alternative and the GDPR. And for the broad comparison on price and features there is ChatGPT alternative: 10 better and cheaper options.

Conclusion