A single request inside GitHub’s Copilot CLI may soon pass through several AI models before any code changes hands. Project HydraFusion, now available as a research preview, decides at runtime how to spend model calls on each task, mixing providers within one workflow instead of locking developers to a single model choice.
The runtime can follow three patterns. A plain single-model path preserves speed for easy jobs. A cascade path starts with a cheaper model and escalates to stronger inference only when a quality gate rejects the output. A critique path sends the first model’s draft to a second model for review, an arrangement GitHub found valuable when an outside perspective beats another unaided attempt. Capability signals for reasoning, code generation, debugging, and tool use guide the planning, and the system prefers the least expensive workflow expected to pass the bar.
Adoption is narrow by design. HydraFusion runs only inside Copilot CLI, works for users on all Copilot plans, and has no open weights or self-hosting option. Enabling it takes three commands: /update, /experimental on, and a /model switch to the HydraFusion entry. Billing follows standard per-token rates for whichever models a workflow invokes. The telemetry stays internal: each step writes its role, result, cost, and latency to a log. What reaches the developer is a single coherent answer and one permission-aware set of changes.
The release extends Auto model selection, which since earlier this year has paired each coding task with a single best-suited model. HydraFusion changes what gets optimized: not the model pick but the plan. To gauge the payoff, GitHub benchmarked fixed policies on CheckpointBench, a replayable internal suite drawn from real Copilot sessions, plus other agentic coding sets, with Claude Opus 5 and GPT-5.6 Sol running at medium reasoning as reference points.
