Quickstart: LiteLLM

One route. Keep LiteLLM for auth, routing, and spend tracking. The race happens beneath it, in your own Airmet sidecar.

# litellm config.yaml
model_list:
  - model_name: qwen3-8b-airmet
    litellm_params:
      model: openai/qwen3-8b
      api_base: http://airmet:8080/v1

Route the models you want raced through Airmet and leave the rest untouched. LiteLLM’s spend log records the leg that answered. The hedge object on each response says how many upstream requests it made; the second leg’s tokens land on your provider invoice and in the sidecar’s local records, not in the hedge object. To keep LiteLLM’s dashboard conservative rather than blind, set the Airmet route’s model_info price to list × your Airmet cost ceiling, or route both Airmet legs back through LiteLLM so its spend log sees every leg. Token spend is not the whole cost on metered upstreams either: each raced request consumes n upstream requests against provider quota; under the default the slower leg is not cancelled, so both legs bill (in full, unless the runaway guard stops one that loops), and under a policy that cuts the loser the cancelled leg still bills its prompt prefill — see the cost note in the OpenAI-compatible quickstart.