Tag
HyDRA is a hybrid dynamic routing architecture for heterogeneous LLM pools that predicts fine-grained capability requirements per query and selects the cheapest capable model via shortfall matching, achieving up to 72.5% cost savings with quality maintained. It is deployed in GitHub Copilot's VS Code Chat auto-mode and decouples routing from model catalog, requiring no retraining when models change.