GitHub opens HydraFusion research preview in Copilot CLI

GitHub announced Project HydraFusion on September 4, 2026, as a research preview for all Copilot plans in Copilot CLI, according to its announcement.

What users can try

Run /update, then /experimental on; select HydraFusion (Research Preview) through /model. Usage is billed at each participating model’s standard token rate.

GitHub says it can use one model, escalate after a quality gate, or add cross-model critique and one revision. GitHub recommends well-scoped, single-prompt tasks.

What the benchmarks establish

GitHub reports that its best-tuned configuration beat Claude Opus 5 by 4.9 percentage points on verified task quality in TerminalBench 2.1 at 67% lower estimated cost. DeepSWE quality was 1.5 points lower at 36% lower cost. These were controlled offline evaluations on benchmarks also used for policy tuning.

How to evaluate it

For a trial, use a fixed repository revision and a small set of representative tasks with acceptance criteria written before execution. Compare HydraFusion with your usual model using the same prompts, tools and permissions. Repeat runs where practical rather than treating one successful patch as a reliable result.

Track test outcomes, reviewer corrections, elapsed time and total token charges per accepted change. A cheaper run that needs more human repair may not lower delivery cost; a more expensive run may still be useful if it reduces rework. Keep conclusions task-specific, and evaluate longer conversations separately rather than extrapolating from a single-prompt trial.