Project HydraFusion: Frontier quality via multi-model orchestration
2 months ago
- GitHub introduces Project HydraFusion, a research preview that uses runtime orchestration to select the best model workflow for each coding task.
- HydraFusion treats workflow selection as an optimization problem, choosing from three execution patterns: single, cascade, and critique.
- It balances quality, cost, and latency by evaluating each request and using the least complex workflow likely to meet its needs.
- In offline evaluations on TerminalBench 2.1, HydraFusion improved task quality by 4.9 percentage points at 67% lower cost compared to Claude Opus 5.
- On DeepSWE and CheckpointBench, HydraFusion achieved near-equivalent quality with 36–65% cost savings.
- HydraFusion is built on five operating principles: execution control, review, cost management, repository state, and permission awareness.
- The research preview focuses on first-turn, single-prompt coding tasks, with future plans for multi-turn performance.
- Developers can select HydraFusion like any other model in GitHub Copilot, and it handles multi-model orchestration behind the scenes.