HydraFusion: Multi-Model Orchestration System
GitHub has added a multi-model orchestration system, HydraFusion, to Copilot. It first evaluates the task and then routes it to the appropriate model. Three modes are available: Single mode for simple tasks handled by one model; Cascade mode where a cheaper model attempts first, escalating to a stronger one if results are insufficient; and Critique mode where one model produces output, a second model from a different family reviews it, and the first model revises accordingly.
Benchmark Performance: On Par with Lower Costs
GitHub compared HydraFusion against Claude Opus 5 on three coding agent benchmarks. It outperformed by 4.9 percentage points on TerminalBench 2.1, lagged by 1.5 points on DeepSWE, and was only 0.1 points lower on CheckpointBench—essentially tied. Costs dropped by 67%, 36%, and 65% respectively.
Comparison with Sakana AI's Fugu
The approach mirrors Sakana AI's Fugu, both turning model selection into multi-model orchestration. The key difference: Fugu is a trained routing model that learns to call different agents and can even recursively invoke itself, while HydraFusion currently selects among three fixed execution modes (Single, Cascade, Critique), effectively baking multi-model scheduling directly into Copilot.
Research Preview and Limitations
HydraFusion is now available as a research preview for all GitHub Copilot plans. It must be enabled via Copilot's experimental features in the CLI. GitHub notes it works best for single-turn coding tasks that can be clearly described; multi-turn long tasks are still being optimized.