GitHub HydraFusion: Copilot CLI orchestrates several models per task
In one sentence On September 4 GitHub opened HydraFusion as a research preview in Copilot CLI: for each request it picks a workflow (Single, Cascade or Critique) and a set of models from different providers. On TerminalBench 2.1 GitHub claims +4.9 quality points and 67% lower estimated cost versus Claude Opus 5.
GitHub is testing a Copilot feature called HydraFusion. Instead of having a single AI model do the work, it decides case by case how to combine several.
For a simple task it uses one model. In other cases it has a cheaper model write a first draft and hands it to a stronger one only if it fails a check. Or it has a second model from another family review the work as a critic, and the first model revises once.
The goal is frontier-level quality at lower cost. For now it is an experimental feature in the command-line version of Copilot, with no service guarantees and not meant for production.
Companies
GitHub, Microsoft
Tools
GitHub Copilot CLI, HydraFusion
Tags
Sources