Auto Router

The Auto Router picks the model for you. Send opendunes/auto as the model; the router classifies the prompt into a task type — debugging code, planning an agent's steps, answering a knowledge question, writing a report — and routes it to the models most spent on for that kind of task over the past week, with fallbacks.

Usage#

curl https://opendunes.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENDUNES_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "opendunes/auto",
    "messages": [{ "role": "user", "content": "Write a SQL query that lists the top 10 wilayas by orders this month." }]
  }'

The answer's model is the model that served; everything else is an ordinary answer. To see the task type the prompt was given, send X-OpenDunes-Metadata: enabled — see Router Metadata.

Settings#

Send them in a plugin: auto-router for opendunes/auto, auto-beta-router for opendunes/auto-beta. Each router reads only its own.

{
  "model": "opendunes/auto",
  "plugins": [
    {
      "id": "auto-router",
      "allowed_models": ["anthropic/*", "openai/gpt-5*"],
      "excluded_models": ["*/claude-haiku-*"],
      "cost_tier": "medium"
    }
  ],
  "messages": [{ "role": "user", "content": "…" }]
}

If your settings leave no model, the request answers 404.

Across a conversation#

The router may pick a different model on each turn. It prefers the model a conversation already landed on while that model stays among the best for the new prompt. Send the same session_id on every turn to mark the conversation.

Billing#

You pay the model that served, at its own price in DA — the router adds nothing. provider.max_price still caps the price of the endpoints it may use.