Audio, embeddings & non-chat endpoints

Auto-routing applies only to chat completions. All other endpoint types bypass the router entirely and reach their real models — even on an auto key. This matters because routing a Whisper request to a text model would fail.

EndpointBehavior
/v1/chat/completionsAuto-routed (on auto keys / router models)
/v1/audio/transcriptionsPass-through → whisper-1 / gpt-4o-transcribe (OpenAI)
/v1/embeddingsPass-through → text-embedding-3-small/-large, ada-002 (OpenAI)
Image generation, moderation, rerank, speechPass-through

Practical effect: an application can use one auto key for everything — chat is cost-optimized by the router, while voice transcription and RAG embeddings just work.