Copilot can already choose a model automatically. But choosing who answers is different from deciding how the work gets done. Project HydraFusion, introduced by GitHub on September 4, tests that second decision in Copilot CLI.

Use the interactive panel above to follow each route:

  • Single: one model delivers the answer.
  • Cascade: a cheaper model tries first; a quality gate decides whether to call a stronger model.
  • Critique: a second model from another family reviews the answer without tool access; the first model then revises it.

The question is no longer just which model to use. It is when another attempt or review is worth its cost.

Cheaper in tests, with trade-offs

The cards show GitHub's comparison with Claude Opus 5. Estimated cost includes every stage, including critique, escalation and retries. DeepSWE captures the trade-off: 36% lower estimated cost, but quality 1.5 percentage points lower.

These are GitHub's own offline tests. CheckpointBench is internal, and the published configuration was the best-tuned one. The figures do not yet establish savings in real development sessions or shorter waits. An extra critique may improve an answer while taking more time.

Try it in the CLI

HydraFusion is a research preview available experimentally across Copilot CLI plans, according to GitHub:

  1. Run /update.
  2. Turn on /experimental on.
  3. Open /model and select HydraFusion (Research Preview).

GitHub recommends substantial, well-scoped work in a single prompt. Billing follows the tokens each model consumes. To assess it as a team, compare a representative task by verified outcome, total cost, elapsed time and human review effort. Long, iterative sessions remain outside the preview's main focus.

Sources: GitHub's research and results, September 4, 2026 and Copilot update, September 10, 2026.