The useful change
Choosing an AI coding model usually happens before the work starts. GitHub’s HydraFusion preview moves part of that choice into the task itself: the system selects a workflow as well as the models that will run it.
GitHub describes three paths. One model can handle a request directly. A cascade can start with an efficient model and escalate when its answer does not clear a quality check. A critique path brings in a model from another family to review the draft before a revision.
Why the critic matters
The interesting distinction is between doing more work and checking the work already done. Asking the same question repeatedly can produce more text without exposing the original mistake. An independent review step has a different job: look for a problem in the proposed result.
That suggests a useful personal comparison. Give a coding assistant a small, well-defined change, then ask it to explain how the change was checked. Look at whether the review found a concrete issue and whether the final patch resolves it.
Trying it
GitHub’s September 4 announcement places HydraFusion in Copilot CLI’s experimental options. Update the CLI, enable experimental features, and select the research preview through the model menu. Usage follows the token charges of the underlying models.
This remains a preview. GitHub’s controlled benchmark results describe its evaluated settings; they do not establish the outcome or cost of your own repository task. The accompanying image is GitHub’s architecture diagram, not a screenshot of our test.
Source published 2026-09-04. Coverage is based on the maker’s announcement and demonstration.