Custom Models (Bring Your Own Key)
Add your own LLM by connecting your provider's API key, and understand which hosting provider and capability flags to choose.
A custom model lets you use a large language model through your own provider account instead of the models DeutschlandGPT ships. You pick a hosting provider (OpenAI, Anthropic, Azure, Google, AWS Bedrock, Mistral, …), paste in the model's ID and your credentials, and the model appears in the model picker like any other.
How your key is handled. Your API key (or service-account credentials) is stored encrypted. When you chat with a custom model, the request goes directly to your provider using your key, so your provider bills you for usage. Because DeutschlandGPT is not the one charging you, the price shown for a custom model inside DeutschlandGPT is always 0.
Choosing a hosting provider
The hosting provider is who runs the model — not necessarily who built it. For example, Anthropic's Claude models can be reached directly through Anthropic, or through AWS Bedrock, or through Google Vertex. Pick the option that matches the account and credentials you already have.
| Provider | When to choose it | Example Model ID | Credentials it needs |
|---|---|---|---|
| OpenAI | You have an OpenAI API key and use the standard Chat Completions API. | gpt-4o | Base URL (default https://api.openai.com/v1) + API key |
| OpenAI (Responses API) | Same key, but you want OpenAI's newer Responses API. | gpt-4o | Base URL + API key |
| Azure OpenAI | Your OpenAI models are hosted in your own Azure resource. | your deployment name | Base URL (your Azure endpoint) + API key |
| Azure OpenAI (Responses API) | Azure-hosted OpenAI models via the Responses API. | your deployment name | Base URL + API key |
| Anthropic | You have an Anthropic API key for Claude. | claude-sonnet-4-5 | Base URL (default https://api.anthropic.com/v1) + API key |
| Anthropic via Google Vertex | You reach Claude through your Google Cloud / Vertex project. | claude-sonnet-4-5 | Google Cloud project + service-account email + region + service-account private key |
| Google Vertex AI | You run Gemini through Google Cloud Vertex. | gemini-2.5-pro | Google Cloud project + service-account email + region + service-account private key |
| Google AI Studio | You have a Google AI Studio API key for Gemini. | gemini-2.5-pro | Base URL (default AI Studio endpoint) + API key |
| AWS Bedrock | You run models through AWS Bedrock with IAM credentials. | anthropic.claude-3-5-sonnet-20241022-v2:0 (Bedrock model ID or full ARN) | Region + access key ID + secret access key |
| AWS Bedrock (Mantle) | OpenAI-compatible Bedrock endpoint authenticated with a Bedrock API key. | Bedrock model ID | Region + API key |
| Mistral | You have a Mistral API key. | mistral-large-latest | Base URL (default https://api.mistral.ai/v1) + API key |
The Model ID must match the identifier your provider expects exactly. For Azure it is the deployment name you created in the Azure portal, not the base model name. For AWS Bedrock it is the Bedrock model ID or the full model ARN. A wrong ID is the most common reason a custom model fails to answer.
For the two Vertex providers (Google Vertex AI and Anthropic via Google Vertex) you also enter your Google Cloud project — the project ID that owns the Vertex endpoint (e.g. my-gcp-project). It sits next to the region and service-account fields and is required for those providers.
Capability flags
Below the credentials, three checkboxes tell DeutschlandGPT what the model can do. They don't change the model — they tell us how to talk to it. If you turn one on that the model doesn't actually support, requests can fail; if you turn one off, we simply won't use that feature.
Tool calls — Lets the model use tools and integrations during a chat (web search, connected services, workflows, and other function calls). Leave it on for modern models. Turn it off only if the upstream model rejects tool definitions, or is a plain text model with no tool support.
Vision (image analysis) — Lets you attach images and have the model read them. Turn it on only if the model actually accepts image input; a text-only model will error on images.
Reasoning — Enables extended thinking / reasoning-effort control for models that support it. For Anthropic models, the correct thinking wire-format is detected automatically from the model ID, so you don't have to configure anything beyond enabling this flag.
Not sure whether a model supports a capability? Check your provider's model documentation. When in doubt, start with Tool calls on, Vision on for multimodal models, and Reasoning off, then adjust if you see errors.
Pricing (optional)
Below the capability flags you can optionally record a cost per 1 million input tokens and cost per 1 million output tokens, in euros. These figures are used only for internal cost attribution inside DeutschlandGPT — usage reports and cost breakdowns. They do not change what you are billed: your provider still bills you directly for the underlying calls. Leave both fields empty if you don't need cost tracking.