Summary
On September 4, 2026, GitHub launched Project HydraFusion as a research preview in the Copilot CLI. It treats model selection as an optimization problem, building a per-task execution plan across models from multiple providers using Single, Cascade, or Critique patterns to hit a quality bar at lower cost. GitHub reports a 67 percent cost reduction and a 4.9-point accuracy gain versus Claude Opus 5 on TerminalBench 2.1.
What changed
GitHub shipped HydraFusion to the Copilot CLI via the /experimental command on September 4, 2026: runtime multi-model orchestration that drafts, critiques, revises, or cascades across providers, choosing Single, Cascade, or Critique execution patterns per task using capability signals for reasoning, code generation, debugging, and tool use.
Why it matters
As frontier model prices and capabilities diverge, routing each sub-task to the most cost-effective model can beat any single model on cost and quality; HydraFusion makes multi-model orchestration a default coding-agent capability rather than something teams hand-build, pressuring rivals to compete on routing, not just model access.
Evidence excerpt
In TerminalBench 2.1, HydraFusion reduced estimated costs by 67 percent compared to Claude Opus 5 while increasing correctly completed tasks by 4.9 percentage points; it chooses Single, Cascade, or Critique execution patterns per task in the Copilot CLI.