GitHub launched Project HydraFusion, a research preview in GitHub Copilot CLI. It plans each coding task at runtime and picks models from several providers. It can use one model alone, escalate a draft to a stronger model, or have a second model review and revise. Developers select it like any other model and see one result. It is open to all Copilot plans via the /experimental command, billed at each model's standard token rate. GitHub tested it offline against Claude Opus 5 and GPT-5.6 Sol on three coding benchmarks. Quality matched or beat Opus 5 at lower estimated cost. Review steps run without tools so they cannot change the repository. No patch is applied if a workflow fails or is cancelled. Intermediate drafts stay hidden until a final answer is ready.
What changed
Copilot's Auto model selection picked one model per task.
What it unlocks
Letting Copilot combine several models per task, with one drafting and another reviewing.
- TerminalBench 2.1: 67% lower cost
- TerminalBench 2.1: +4.9 points quality
- DeepSWE: 36% lower cost, -1.5 points
- CheckpointBench: 65% lower cost
What you need to act on it
- a GitHub Copilot plan
- GitHub Copilot CLI updated to the latest version
- /experimental turned on
- token usage billed at each model's standard rate
Sources