GitHub stops picking a model and starts building the workflow: HydraFusion cuts cost by up to two thirds, quality about level
HydraFusion decides at runtime which models run and in what order for each request. Across three benchmarks cost fell between 36% and 67%; quality rose 4.9 points on one and barely moved on the other two.
GitHub has opened Project HydraFusion for Copilot as a research preview. The idea described in the post dated 4 September goes past model choice: for each request, the system constructs at runtime which models run and in what order.
Three execution patterns
When a request comes in, HydraFusion picks one of three shapes:
- Single: one model solves the task directly.
- Cascade: an efficient model drafts a solution; a quality gate decides whether to accept it or escalate to a stronger model.
- Critique: one model drafts, an independent critic from a different model family reviews it, and the drafting model revises once.
The system picks the least complex workflow expected to meet the need, using additional model calls only when they are likely to improve the result.
Three benchmarks, three different answers
GitHub measured against Claude Opus 5 and GPT-5.6 Sol. All three results:
| Benchmark | Cost vs Opus 5 | Quality |
|---|---|---|
| TerminalBench 2.1 | 67% lower | +4.9 points |
| DeepSWE | 36% lower | −1.5 points |
| CheckpointBench | 65% lower | −0.1 points |
So the real result is not "better" but "about the same quality, markedly cheaper". GitHub's own wording points the same way: "In offline evaluations across three agentic coding benchmarks, HydraFusion consistently demonstrated frontier-level quality with substantial estimated cost savings."
CheckpointBench is GitHub's internal benchmark based on real Copilot sessions. The company also notes in its own post that TerminalBench 2.1 is relatively saturated, which makes broader validation important.
The limits the company states itself
- The system currently suits first-turn, single-prompt coding tasks; multi-turn performance work is forthcoming.
- Progress visibility is limited: intermediate drafts are held until the final result.
- The post carries this warning verbatim: "Results, models, workflows, availability, names, and product behavior may change as we learn from the preview."
How to try it
HydraFusion is available inside GitHub Copilot CLI through /experimental, on all Copilot plans. Pricing is based on the tokens consumed by the models it uses, at each model's standard rate.
Why it matters
On the enterprise side, model choice has been lived as a fixed decision: a team picks a model and runs everything through it. HydraFusion takes that decision per request, and that is where the cost comes down. Simple work stays on the cheap model, hard work escalates to the expensive one, contested work gets criticized by a second model family.
The line for a corporate buyer is not the quality column but the 36% to 67% drop in cost. And the condition under which that was measured is offline evaluation; whether the gain survives live repositories, tool failures and human review has not been shown yet.
