AI 程式開發工具
GitHub HydraFusion Orchestrates Multiple Models at Copilot Runtime, Cutting Costs but Not Consistently Improving Quality
Rather than relying on the same model for every step, HydraFusion chooses among a single model, cascading escalation, and cross-model review based on the task. GitHub’s offline tests showed lower costs across all three benchmarks, but only TerminalBench 2.1 also delivered higher quality.
GitHub has added Project HydraFusion to Copilot CLI as a research preview, elevating model selection into a runtime decision spanning the entire workflow. Users simply select HydraFusion, and the system routes work among three modes based on capability signals for reasoning, code generation, debugging, and tool use: `Single`, in which one model completes the task directly; `Cascade`, in which a less expensive model produces an initial draft before a quality gate determines whether to escalate; and `Critique`, in which a read-only model from another model family reviews the output before the original model revises it once.
The engineering focus of this design extends beyond the router itself. GitHub aggregates all usage from drafting, review, retries, escalation, and fallback; sets timeout and cancellation rules for each stage; and runs the reviewing model in an isolated environment without tools. Only the solving model can access the shared workspace, where permissions are controlled. If validation fails or the workflow is canceled, the system applies no patches and records each step’s role, cost, latency, and diagnostic data.
Compared with Claude Opus 5, the best-tuned configuration improved verified task quality on TerminalBench 2.1 by 4.9 percentage points while reducing estimated cost by 67%. On DeepSWE, however, cost fell by 36% while quality also declined by 1.5 points. On the internal CheckpointBench, it saved 65% but scored 0.1 points lower on quality. These results suggest a configurable quality–cost trade-off rather than an inherent accuracy advantage from using multiple models.
Engineering teams should still pay attention to real-world latency, routing observability, and strategy drift after model updates. The results come from offline evaluations using a fixed model pool, and CheckpointBench has not been made public. For now, GitHub recommends HydraFusion primarily for clearly scoped, single-turn tasks; its effectiveness in long-running, multi-turn development workflows remains to be validated.