GitHub announced the GitHub HydraFusion preview on September 4, 2026. The research feature coordinates models from multiple providers within a coding task and is available across Copilot plans through experimental mode in Copilot CLI. Source: GitHub’s announcement.
What changed
HydraFusion selects between a single-model response, a draft that can escalate to a stronger model, and a draft reviewed by a separate model family before revision. GitHub says usage is charged for the participating models’ tokens at their standard rates.
To access it, GitHub instructs CLI users to run /update, enable /experimental on, and select HydraFusion (Research Preview) through /model. The company recommends substantial, well-scoped, single-prompt coding tasks for the preview; longer iterative sessions remain an area of development.
GitHub reports favorable quality and estimated cost results in controlled offline coding evaluations. These are vendor benchmark findings, not evidence of equivalent gains in your test suite. Availability and behavior may change during research.
Why this matters for QA engineers
QATechTools recommendation: evaluate the complete workflow against a fixed baseline before using it for regression repairs. A second model’s review is useful evidence to inspect, but your executable acceptance checks should still decide whether a patch passes.
- Use known defects: require each generated regression test to fail on the original bug and pass after the fix.
- Measure total effort: record end-to-end latency, total usage, and human review time across repeated runs.
- Inspect the final diff: reject repairs that weaken assertions, skip failing tests, or change unrelated behavior.
Start with one bounded defect reproduction and repair task. Compare its independently verified outcome with your existing assistant workflow before expanding adoption.