Chat

DONNA

Chat endpoint with reasoning, streaming, and tool calling. DONNA Cadenza is a general-purpose assistant, DONNA Coda specialises in code.

Models

All models speak the same API and differ in depth, limits and price.

donna-cadenzaGeneral-purpose assistant

The default for almost everything: knowledge questions, writing, summaries, translation, analysis. Use this unless you have a reason not to.

Price per 1,000 tokens
40 / 200 credits (input / output)
Context window
256.000 tokens
Max output
128.000 tokens
Thinking
configurable, can be switched off
Image input
yes
Tool calling
yes
donna-codaCoding specialist

Same price and same limits as donna-cadenza, but tuned for software work — writing code, debugging, refactoring.

Price per 1,000 tokens
40 / 200 credits (input / output)
Context window
256.000 tokens
Max output
128.000 tokens
Thinking
configurable, can be switched off
Image input
yes
Tool calling
yes
donna-sinfoniaFlagship for complex work

Built for multi-step analysis, very long documents and tasks that chain several steps. Five times the price of donna-cadenza — not worth it for routine requests.

Price per 1,000 tokens
80 / 400 credits (input / output)
Context window
1.000.000 tokens
Max output
128.000 tokens
Thinking
always on, depth configurable
Image input
yes
Tool calling
yes

Pricing

donna-cadenza
Input per 1,000 tokens
40Credits
Output per 1,000 tokens
200Credits
donna-coda
Input per 1,000 tokens
40Credits
Output per 1,000 tokens
200Credits
donna-sinfonia
Input per 1,000 tokens
80Credits
Output per 1,000 tokens
400Credits
  • Reasoning tokens (thinking phase) count as output tokens and are billed at the output rate.
  • On donna-sinfonia reasoning tokens are incurred on every request — the model is built for multi-step work.

Thinking and cost

DONNA thinks before answering. That thinking phase produces reasoning tokens, billed as output, and they can outweigh the visible answer several times over. Use reasoning_effort to steer the depth: "low" for short factual lookups, "high" for hard problems.

donna-sinfonia is built for complex work and therefore reasons on every request. reasoning_effort steers the depth there but does not switch the thinking phase off — reasoning tokens are incurred either way. If you need an answer with practically no thinking, use donna-cadenza with reasoning_effort: "low".

Context window and output length

Input and output share the same window. If max_tokens is omitted, the gateway picks a budget that fits whatever space is left — so a long conversation cannot overbook the window.

GET/v1/models

Lists the models your key may use — machine-readable, with type and capabilities per model. The model cards above are rendered from exactly these fields. Add ?type=chat or ?type=embedding (comma-separated for several) to narrow the list to one type; without the parameter every model is returned.

Request

curl "https://api.centra.dbpg.io/v1/models" \
  -H "Authorization: Bearer sk_your_api_key"

Response

{
  "object": "list",
  "data": [
    {
      "id": "donna-cadenza",
      "object": "model",
      "type": "chat",
      "created": 1742083200,
      "owned_by": "dbpg",
      "capabilities": {
        "context_window": 256000,
        "max_output_tokens": 128000,
        "reasoning": "configurable",
        "vision": true,
        "tools": true
      }
    },
    {
      "id": "donna-coda",
      "object": "model",
      "type": "chat",
      "created": 1742083200,
      "owned_by": "dbpg",
      "capabilities": {
        "context_window": 256000,
        "max_output_tokens": 128000,
        "reasoning": "configurable",
        "vision": true,
        "tools": true
      }
    },
    {
      "id": "donna-sinfonia",
      "object": "model",
      "type": "chat",
      "created": 1742083200,
      "owned_by": "dbpg",
      "capabilities": {
        "context_window": 1000000,
        "max_output_tokens": 128000,
        "reasoning": "always",
        "vision": true,
        "tools": true
      }
    },
    {
      "id": "donna-chiodi-v20260320",
      "object": "model",
      "type": "embedding",
      "created": 1710892800,
      "owned_by": "dbpg",
      "capabilities": {
        "context_window": 512,
        "vision": false,
        "tools": false
      }
    }
  ]
}
Read capabilities, don't assume them. type separates chat from embedding models — treat any other value as not chattable, so a future type cannot silently appear in a model picker. context_window is what you size compaction or truncation thresholds from; windows differ substantially per model. reasoning is "configurable" when reasoning_effort steers the depth and "always" when the model reasons regardless — a UI must not claim otherwise in that case. vision and tools state whether image parts in the content array, and tool definitions, are accepted. Models under evaluation do not appear in the list but can be called by name.
POST/v1/chat/completions

Chat endpoint with model selection via the model field. Authentication via Bearer token in the Authorization header. Besides the fields listed, the OpenAI standard fields top_p, stop, seed, presence_penalty, frequency_penalty, parallel_tool_calls, n (1 only) and user are accepted. Any other parameter is rejected with 400 — nothing is silently ignored.

Parameter

  • modelstring
    Yes

    Model id: 'donna-cadenza', 'donna-coda' or 'donna-sinfonia'.

  • messagesarray
    Yes

    Conversation history as an array of message objects with the fields role and content. Roles: 'system', 'user', 'assistant', 'tool'.

  • max_tokensintegerOptional

    Upper bound on output tokens, thinking included. Alias: max_completion_tokens. If omitted, the gateway picks a budget that fits the context window.

  • temperaturenumberOptional

    Sampling temperature (0–2). 0 = deterministic.

  • streambooleanOptional

    If true, the response is delivered as Server-Sent Events (SSE).

  • stream_optionsobjectOptional

    Optional, object with field include_usage. When set to true, a final usage chunk is emitted in the stream.

  • reasoning_effortstringOptional

    Thinking-phase depth: 'low' | 'medium' | 'high'. Any other value is rejected with 400. On donna-sinfonia the value steers depth only — the thinking phase does not go away there.

  • toolsarrayOptional

    Tool definitions for function calling.

  • tool_choicestring | objectOptional

    'auto' | 'none' | object with fields type='function' and function.name — controls whether/which tool is called.

Request

curl -X POST "https://api.centra.dbpg.io/v1/chat/completions" \
  -H "Authorization: Bearer sk_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "donna-cadenza",
    "messages": [
      { "role": "user", "content": "Was ist die Hauptstadt von Frankreich?" }
    ],
    "reasoning_effort": "low"
  }'

Response

{
  "id": "donna-7c4e5f8a91b3d2e1",
  "object": "chat.completion",
  "created": 1778407200,
  "model": "donna-cadenza",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Paris.",
        "reasoning_content": "The user is asking about the capital of France. That is Paris."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 32,
    "completion_tokens": 18,
    "total_tokens": 50
  }
}
POST/v1/chat/completions

Vision: send an image as a base64 data URI in the content array (no remote URLs).

Request

curl -X POST "https://api.centra.dbpg.io/v1/chat/completions" \
  -H "Authorization: Bearer sk_your_api_key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "donna-cadenza",
    "messages": [
      { "role": "user", "content": [
        { "type": "text", "text": "Erkläre, was auf dem Bild zu sehen ist." },
        { "type": "image_url", "image_url": { "url": "data:image/jpeg;base64,/9j/4AAQSkZJRgABAQAA..." } }
      ] }
    ]
  }'

Response

{
  "id": "donna-9f2a7c4e5f8a91b3",
  "object": "chat.completion",
  "model": "donna-cadenza",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "Auf dem Bild ist … zu sehen."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": { "prompt_tokens": 1187, "completion_tokens": 94, "total_tokens": 1281 }
}

AI labelling (EU AI Act)

Under Art. 50(2) EU AI Act, every answer of the DONNA chat models is labelled as AI-generated in machine-readable form. Each response carries the header X-Donna-AI-Generated: true and a signed provenance object, when streaming in the last chunk (the one with usage). The field is a dbpg extension; OpenAI-compatible SDKs ignore it.

{
  "provenance": {
    "version": 1,
    "request_id": "donna-8f30334ed6aa46fe",
    "timestamp": "2026-08-04T12:00:00Z",
    "model": "donna-cadenza",
    "provider": "dbpg-centra",
    "output_hash": "sha256:d3751d33f9cd5049c4af2b462735457e4d3baf130bcbb87f389e349fbaeb20b9",
    "signature": "v1:ed25519:MEUCIQ…",
    "public_key_url": "https://api.centra.dbpg.io/.well-known/donna-provenance.json"
  }
}

output_hash is SHA-256 over the concatenated assistant text of all choices (when streaming: all delta.content joined). The signature is Ed25519 over the canonical JSON string with exactly this field order and no whitespace:

{"model":…,"output_hash":…,"request_id":…,"timestamp":…,"version":1}

The public keys are published at public_key_url. Old key versions stay available there permanently, so signatures remain verifiable after a key change.

POST/v1/provenance/verify

Checks a certificate against the text it covers. Free and without an API key, so third parties can verify too; at most 1 MB per request and 60 requests per minute per IP address. On failure, reason names the cause: output_hash_mismatch (text does not match the certificate), signature_invalid (certificate altered or not ours), unknown_key_version, malformed_signature or malformed_certificate.

Request

curl -X POST https://api.centra.dbpg.io/v1/provenance/verify \
  -H "Content-Type: application/json" \
  -d '{
    "provenance": { … das provenance-Objekt aus der Response … },
    "output_text": "Die Antwort ist 42."
  }'

Response

{ "valid": true }
Legal. Terms, privacy and additional documentation for the DONNA model family at buehnenproduktionsgesellschaft.de/donna.