GitHub has introduced Project HydraFusion, a Copilot research preview that assembles different AI models into a workflow for each coding request. Developers select HydraFusion like a normal model; the service decides whether to use one model or coordinate several behind the scenes.
The router currently has three patterns. “Single” sends straightforward work to one model. “Cascade” lets an efficient model try first, then escalates when an acceptance check fails. “Critique” gives a second model the role of reviewing a draft before revision. GitHub says the system uses signals related to reasoning, code generation, debugging and tool use to choose the least complex route expected to meet its quality target.
In controlled offline tests on three agentic coding benchmarks, one tuned configuration improved verified task quality on TerminalBench 2.1 by 4.9 percentage points while using 67 percent less estimated cost than Claude Opus 5. Cost accounting included every draft, critique, retry and fallback call.
Those results are limited to the tested benchmark versions, model pool, pricing assumptions and medium reasoning setting. The preview will provide a more realistic test of latency and repository-level behavior. GitHub records each model’s role, result, cost and diagnostics internally, while presenting the developer with one response and one permission-aware set of changes.