GitHub has introduced Project HydraFusion, a research preview for Copilot that treats model choice as a runtime orchestration problem rather than a single dropdown selection.

What changed

The preview lets users select HydraFusion like a model in Copilot. Behind the scenes, GitHub says it builds an execution plan that can pick models from multiple providers, draft an answer, run a critique-and-revision pass, or escalate to a stronger model when the task appears to need it.

GitHub describes three current patterns: direct solving by one selected model, a cheaper first attempt with a quality gate before escalation, and a draft-review-revise loop using an independent critic from a different model family. Review steps run in isolated, tool-less contexts, while solver steps use the normal permission-aware agent loop.

Why it matters

The claim is not simply that another model is available. GitHub is testing whether compound coding workflows can preserve frontier-level quality while reducing estimated cost for agentic tasks. In its offline evaluation, the company says HydraFusion improved verified task quality on TerminalBench 2.1 by 4.9 percentage points at 67% lower estimated cost than Claude Opus 5. It also reported lower estimated cost on DeepSWE and its internal CheckpointBench, with results tied to the benchmark versions and pricing assumptions used.

For developers, the conservative takeaway is that coding assistants are becoming orchestration systems. The visible interface may still look like one Copilot response, but the product is increasingly deciding which model, workflow, and review path should handle each request.