Configuring language models

Purpose

dAI Pro talks to several providers of language models. An LLM configuration bundles provider, access, models and region; collections pick one configuration each for chat and for embeddings from it, plus additional models for comparisons. This way the provider can be changed without touching a collection, and each collection can have its own access with its own billing.

Where

LLM configurations in the top bar. The list shows all configurations with type, models and region; the order determines the selection lists. "Add configuration" asks for the provider type first.

The providers

Type	Chat	Embeddings	Audio	Images	Region
Mistral AI	yes	yes	yes (Voxtral)	no	EU
OpenAI	yes	yes	yes	yes	per contract, usually USA
Anthropic (Claude)	yes	no	no	no	USA
Google Gemini	yes	yes	no	yes	per contract
OpenAI-compatible endpoint	yes	depends on the server	no	no	freely selectable

Default models. Every configuration comes with a preset chat model and preset additional models; empty fields mean this default, which is shown as a placeholder in the field. As of 24 September 2026, prices in USD per million tokens (input / output):

Provider	Chat model (default)	Additional models (default)	Why
Mistral AI	`mistral-large-latest` (Large 3, 0.5 / 1.5)	`mistral-medium-latest`, `mistral-small-latest`	Open flagship at a fraction of the competition's price. Mistral itself recommends `mistral-medium-latest` (Medium 3.5, 1.5 / 7.5) for agents and code: stronger, three times the price.
OpenAI	`gpt-6-sol` (2 / 10)	`gpt-6-luna` (0.10 / 0.50)	Successor of gpt-5.5 at a third of the price; `gpt-6-astra` (10 / 50) only for the hardest tasks, no tool calls.
Anthropic (Claude)	`claude-sonnet-5` (2 / 10)	`claude-haiku-4-5` (1 / 5)	Balanced; `claude-opus-5-5` (4 / 20) for hard tasks.
Google Gemini	`gemini-3.8-flash` (0.75 / 3.75 until end of 2026, then double)	`gemini-3.5-flash-lite` (0.30 / 2.50)	Newest Flash model at the price of 3.6 and 3.7; Google recommends it for new projects.
IONOS (preset)	`openai/gpt-oss-120b`	`meta-llama/Llama-3.3-70B-Instruct`	Open reasoning model hosted in Germany.
STACKIT (preset)	`openai/gpt-oss-120b`	`openai/gpt-oss-20b`	Same model, region Baden-Württemberg.
Mittwald (preset)	`gpt-oss-120b`	`Ministral-3-14B-Instruct-2512`	Mittwald's recommendation for cost-conscious chat as the second model.
OVHcloud (preset)	`gpt-oss-120b`	`Mistral-Small-3.2-24B-Instruct-2506`	From the AI Endpoints catalog.
Scaleway (preset)	`meta/llama-3.3-70b-instruct:fp8`	—	Catalog keeps growing (GLM 5.2, Qwen 3.6); check the identifiers in the console.
Ollama, vLLM (preset)	—	—	Your server determines the model; enter it.

Model identifiers change at every provider every few months. What your key really sees is given by the provider's model list (GET /v1/models on all OpenAI-compatible endpoints).

For images the two providers differ. OpenAI delivers transparent backgrounds and free pixel sizes. Gemini keeps the selected aspect ratio, for instance 16:9 at 1376 × 768 pixels in the 1K tier. dAI Pro chooses the tier 1K, 2K or 4K by the size of the requested image, with the quality “High” at least 2K. OpenAI delivers at most 3,840 pixels on the long side, Gemini up to about 5,500 pixels in the 4K tier; dAI Pro enlarges larger formats afterwards (Chapter 18, section “Formats and resolution”). Gemini does not deliver transparent backgrounds. Both accept reference images from the corporate design as a style guide. For illustrations with an exact background or cut-out images you need OpenAI (Chapter 18).

The type "OpenAI-compatible endpoint" covers every server that speaks the OpenAI dialect: European providers such as IONOS, STACKIT, Mittwald, OVHcloud and Scaleway, but also your own server with Ollama or vLLM. For the well-known providers there are presets that fill in endpoint, models and region.

Costs and prices

You enter the list prices per model under "Usage" (Chapter 19); they apply to all collections. For image models the price table also has the column “€/image”. If the image model reports tokens and token prices are stored, dAI Pro calculates exactly by tokens; the price per image applies only to image models that report no tokens. Never both. OpenAI and Gemini report tokens. With OpenAI, an image in the quality “High” costs about six times as much as “Standard”; example prices are in Chapter 19. The provider's billing runs per API key, which is why a separate key per customer pays off.

The operator’s starter quota

If your instance is hosted by the operator (managed) and a starter quota was agreed on the order sheet, the instance initially answers via the operator’s Mistral key — up to a value of 10.00 €. The quota is a one-off and ends when it is used up:

From 80 % on you see a yellow notice on every page and the administrators receive an e-mail. Then create a configuration here with your own Mistral API key and select it in your collections — calls with your key do not count against the quota, and billing runs through your Mistral account.
At 100 % the operator’s key is disabled for your instance: chat, search, indexing and agents use the language model again only once your own key is in place. Until then visitors see “The assistant is temporarily unavailable” in the chat.

The quota is calculated with the prices under “Usage”; the usage page shows which calls counted against it.

Frequently asked questions

The chat answers with an "authentication" error. The API key has expired, has been blocked, or the credit is used up. Check the account at the provider; a new key is simply entered in the configuration.

Can I delete configurations? Only if no collection uses them. The list shows the usage.

Which model is the right one? For most websites, a current large model of the provider for the chat and the standard embedding model. Compare via "Additional language models" in the test chat before you switch.