GitHub Releases HydraFusion: 67% Cost Reduction via Multi-Model Orchestration

GitHub released Project HydraFusion as a research preview. This project dynamically orchestrates models from multiple providers at runtime, offering a workflow that balances performance, cost, and latency while allowing developers to use it as if selecting a single model.

HydraFusion evaluates each request and automatically selects the most efficient execution pattern. The key patterns are as follows:

• Single: A single model directly solves the task
• Cascade: An efficient model drafts the response, escalating to a more powerful model if it fails to pass quality gates
• Critique: One model drafts the response, and another model independently reviews and revises it

Offline evaluation results showed that on the TerminalBench 2.1 benchmark, HydraFusion improved verified task quality by 4.9%p compared to Claude Opus 5 while reducing estimated costs by 67%. It also achieved cost savings of 36% and 65% on DeepSWE and CheckpointBench, respectively, while maintaining frontier-level quality.

The system is built on five operational principles, including Isolated review, Fail-safe application, and Validated routing, strengthening repository state control and cost management. It is currently available to all plan users via the /experimental command in GitHub Copilot CLI, with usage fees based on actual model tokens consumed.