AI coding tools
GitHub HydraFusion Orchestrates Multiple Models at Copilot Runtime, Using Cascades and Cross-Review to Control Costs
The Copilot CLI research preview no longer selects just one model for each task. Instead, it routes tasks among three workflows: direct execution, staged escalation, and cross-model review. GitHub’s offline tests indicate that some workloads may reduce estimated costs, but the router, quality gates, and internal benchmark have not been disclosed.
GitHub launched Project HydraFusion on September 4, turning model selection from a one-time decision made before a prompt is submitted into a composite runtime workflow. Users select it as a model in Copilot CLI, while the system chooses one of three paths based on signals related to reasoning, code generation, debugging, and tool-use capabilities: `Single` sends the task directly to one model; `Cascade` has a more resource-efficient model generate a solution first and escalates only if it fails a quality gate; and `Critique` has a different model family review the draft in a read-only, tool-free environment before the original model revises it once.
This isolation model is notable. The problem-solving models share the workspace and operate under Copilot’s permission controls, while the review model cannot modify the repository. If the workflow is canceled or validation fails, the system does not apply any patch. The executor records each stage’s role, outcome, latency, diagnostics, and cost. Billing covers all tokens consumed by drafting, review, retries, escalation, and fallback. Calling multiple models is therefore not necessarily cheaper; savings depend on whether the router can keep enough tasks on lower-cost paths.
GitHub tested fixed strategies on TerminalBench 2.1, DeepSWE, and its internal CheckpointBench. Compared with Claude Opus 5, the best-tuned configuration delivered 4.9 percentage points higher validated quality on TerminalBench 2.1 at 67% lower estimated cost. On DeepSWE, it was 36% cheaper but scored 1.5 percentage points lower in quality. On CheckpointBench, cost was 65% lower and quality was 0.1 percentage points lower. These are GitHub-controlled offline results using the best-performing configurations. The company has not disclosed the routing model, quality gates, per-task distribution, or end-to-end latency, and the internal benchmark cannot be independently reproduced.
The preview is available across all Copilot plans, but GitHub recommends initially using it for well-scoped, first-pass, single-prompt tasks; longer, multi-turn sessions are still under development. Engineering teams should next monitor tail latency in real repositories, failure fallbacks, cost predictability, and whether enough routing logs can be exported for auditing. If these controls remain opaque, the model-orchestration layer could become a new source of vendor dependence and a debugging blind spot.