HydraFusion

Traditional model selection picks one model per request. HydraFusion goes further: it treats each coding task as an optimization problem and builds a workflow around it. For simple tasks it routes to a single efficient model. For harder ones it applies a cascade pattern, where a smaller model drafts and a quality gate decides whether to escalate to a stronger model. For the hardest tasks it uses a critique pattern, where one model drafts and a separate model from a different family reviews and flags issues before the drafter revises.

GitHub announced HydraFusion on September 4, 2026 as a research preview in GitHub Copilot CLI, available to all plan holders via the /experimental flag. In benchmark testing against running a single frontier model end-to-end, it cut estimated costs by 67% while improving task completion rates by nearly five percentage points. Billing is standard: you pay the token rates for whichever models the orchestrator actually invokes, with cheaper models handling most of the work.

The broader significance is conceptual: HydraFusion is an early mainstream example of 'model orchestration' replacing 'model selection' as the default mental model. Instead of asking 'which model should I use?' the system asks 'what workflow should I run?' That framing is spreading across agent harnesses and coding tools, and is a useful lens for builders thinking about cost and quality tradeoffs in production.