Building an AI chatbot in Dutch: what are your options?

Building a Dutch-language AI chatbot is no longer a technical problem. Every major provider has an API, you connect a few hundred lines of code to it and you have something working within a day. The questions that actually matter are different: which model writes Dutch you'd be comfortable showing your customers, what that costs per conversation, and where your users' questions end up.

This article covers the three routes you have — a commercial API, an open model you run yourself, or a combination — with token prices from August 2026 and a calculation you can check yourself. It isn't about which framework to use, because that changes every quarter. It's about the choices that determine your monthly bill and your Dutch output.

Three routes, three kinds of work

Route 1: a commercial API. You send your conversations to OpenAI, Anthropic, Google, xAI or Mistral and pay per million tokens. You have no infrastructure, you get the best models and you can be live within a day. The price scales linearly with usage, and your data leaves your own environment.

Route 2: an open model you run yourself. You download the weights, put them on a server of your choice and pay for compute instead of per token. You have full control over where the data lives, but you also take on the work of scaling, updates and availability yourself.

Route 3: a hybrid. Most serious chatbots end up doing this: a cheap model for routine questions and classification, an expensive model only for the conversations where it really matters. That cuts costs by a factor of ten without users noticing.

What the APIs cost per million tokens

The prices below are in dollars per million tokens, split into input and output. The last column is the amount for 10,000 replies, calculated with a realistic average of 1,500 input tokens per turn (system prompt, knowledge snippets and the question) and 500 output tokens.

The interesting part is the ratio, not the absolute amounts. There's a factor of 76 between the most expensive and the cheapest model in this table. For a chatbot that answers frequently asked questions, that's the difference between a minor line item and a cost you need to discuss with your client.

Also keep in mind that these prices are in dollars and exclude VAT. At an exchange rate of $1 = €0.8687, $60 is about €52.12 excluding VAT. If you have a VAT number, that's the amount that sticks; as a private individual, 21% is added on top.

Which models write good enough Dutch

Price is half the story. The other half is whether the output looks like it was written by a Dutch speaker. In our test of Dutch writing quality per provider, Claude is the only one to score a 5, with ChatGPT, Gemini, Perplexity and Le Chat on a 4.

For a chatbot, that means the following. If you use a cheap model for volume, expect the tendency toward anglicisms and literally translated expressions to increase as the model gets smaller. You can partly fix that with your system prompt: fix the form of address, give two or three example replies in the tone you want, and include a short list of words the bot must never use.

What you can't fix with a prompt is a feel for colloquial Dutch. If your bot runs on a channel where people write informally, a stronger model is worth the investment on that point. An overview of all models and what they're meant for is in all AI models side by side.

Open models you can run yourself

Self-hosting has become more realistic over the past year, mainly because the licenses are looser.

Mistral Small 4 and Mistral Large 3 are released under Apache 2.0, which means you can use them commercially without additional conditions. Small 4 is multimodal and has 256K context; Large 3 is larger and dates from December 2025. Mistral Medium 3.5 has open weights under a modified MIT license.

Meta's Muse Glimmer is the most interesting newcomer for smaller projects: 30B parameters, Apache 2.0, 256K context, and it runs on a single consumer GPU. That makes local testing or a small internal bot feasible without a data center.

The Llama 4 models, Scout and Maverick, are still the standard for self-hosting, with context windows of up to 10M tokens. There are also open models with a clear specialty: GLM-5.2 from Z.ai for engineering work, Kimi K2.7 Code from Moonshot for long coding-agent runs, and Qwen 3.8 Max from Alibaba as the newest model in that corner.

The trade-off is simple to make. Self-hosting pays off as soon as your volume is high and predictable, or as soon as you have strict requirements about where data is stored. With low or erratic volumes, your own infrastructure means paying for capacity that sits idle, whereas with an API you only pay for what you use.

Hosting and processing in the EU

If you're building a chatbot for Dutch users, the question of where conversations are processed is almost always part of the discussion with your client. Three things are firm here.

Mistral processes in the EU, in France and Sweden. That makes it the shortest route to European processing without having to host anything yourself. Microsoft has an EU Data Boundary for business customers, which is relevant if your bot lands in a Microsoft environment. And DeepSeek runs in China without an adequacy decision, which makes its models via the official API unsuitable for personal data or customer data, however attractive the price of $0.14 per million input tokens may be.

If you run open weights on your own infrastructure, you decide which region your server is in and which data leaves it. That's the strongest argument for route 2 in sensitive applications.

What it costs in practice

A customer service bot with 300 conversations a day and an average of four turns per conversation comes to around 36,000 replies a month. With Mistral Small 4 that costs about $19 a month, with Claude Sonnet 5 about $288, and with GPT-5.6 Sol about $810.

That difference explains why route 3 is so popular. Let a cheap model work out what the question is about and handle the standard questions, and only pass the complex conversations on to an expensive model. In practice, most questions are about the same handful of topics, so you get the costs of the cheap column with the quality of the expensive one.

If you're building this for a client, there's one more cost that's often forgotten: your own subscriptions for developing and testing. What that costs, and what you can do with it for tax purposes as a Dutch freelancer (zzp'er), is covered in AI for freelancers and small businesses and in our comparison of AI subscriptions.

If you're building something where European processing is the starting point, you can read the broader background of that choice in why Dutch users are switching to European alternatives.

Conclusion