GitHub: Copilot HyDRA Router Automatically Selects Model and Delivers 72.5% Savings at Same Quality

GitHub described HyDRA, a model routing system that enables Copilot to automatically select the most appropriate AI model for each task, reducing cost without sacrificing quality.

Routing by Task Complexity

HyDRA is a model router that selects the AI model based on four criteria: reasoning depth, code complexity, debugging difficulty, and tool orchestration requirements. Instead of routing every query to the same large model, simple tasks receive a cheaper model while complex ones get a more capable one. This avoids spending expensive reasoning capacity on trivial queries.

Measurable Savings Versus Competition

GitHub reports 72.5% cost savings while maintaining quality. In conservative mode, HyDRA achieves a 70.8% task resolution rate with 3.3x lower cost than OpenRouter Auto. This is complemented by prompt caching, which reuses model state for repeated prefixes to reduce redundant processing, and Tool Search, which fetches tool definitions on demand rather than loading all of them upfront.

How Does It Perform in Other Languages?

Routing accuracy remains within 4 percentage points of the English baseline across 19 languages and 16 language families, meaning multilingual users do not lose out on model selection quality. Auto mode only activates at natural cache boundaries (first turn or after compaction), preserving cache efficiency.