GitHub Launches HydraFusion, Citing 67% Lower Cost on Terminal-Bench 2.1
The Copilot research preview turns model selection into a per-task workflow. Its strongest published savings result is a GitHub-run, tuned offline evaluation—not a measure of ordinary repository work.
Listen to this story
The audio brief
Story brief
3 key pointsHydraFusion is GitHub’s experimental Copilot CLI router for sending coding tasks through a single model, an escalation cascade, or a draft-and-critique workflow. GitHub reports near-Claude Opus 5 benchmark quality at 36–67% lower estimated cost, but those results come from tuned offline policies and vary by benchmark. Developers still pay standard token rates for every model invoked, with no orchestration surcharge....
- 01
HydraFusion offers single-model, escalation, and cross-model critique routes; reviewers cannot use tools or modify repositories.
- 02
GitHub reports 4.9-point higher Terminal-Bench 2.1 performance than Claude Opus 5 at 67% lower estimated cost.
- 03
On DeepSWE, HydraFusion trails Claude Opus 5 by 1.5 points while reducing estimated cost 36%.
GitHub has launched Project HydraFusion, a Copilot research preview that can send a coding task through several models rather than hand the whole job to one. The practical bet is that routing can reserve more expensive inference for requests that need it.
HydraFusion is available through the experimental menu in GitHub Copilot CLI on every Copilot plan. Developers select it from the model picker, and the service chooses the workflow and models behind the request before returning one response and proposed change set.
Three routes for one coding request
The router has three options. A single-model workflow sends the task directly to one selected model. A cascade begins with an efficient model and calls a stronger one only if an acceptance check rejects the first result. In a critique workflow, a separate reviewer from a different model family reads a draft before its original author gets one revision.
That review is separated from the agent that can act on code. Review models have no tools and cannot modify the repository; solver models retain the shared workspace and Copilot’s permission-aware agent loop.
Savings depend on the route holding up
The three results point in different quality directions, even as each shows a lower estimated cost. GitHub describes them as offline evaluations using selected benchmark revisions and tuned routing policies. It tuned policies with beam search over per-capability scores across the three benchmarks, so the published figures reflect tuned offline policies.
The preview’s next test is sustained work
A multi-model route does not make extra inference free. Usage is billed for tokens consumed by every model in the workflow at each model’s standard Copilot rate, although GitHub says it adds no orchestration surcharge. A critique or escalation therefore has to earn back its additional calls through cheaper routing elsewhere.
GitHub recommends first-turn, single-prompt coding tasks as the preview’s current use case. Multi-turn sessions remain a development focus, and HydraFusion withholds intermediate drafts because later review or escalation may discard them. The unresolved question is whether its benchmark-tuned trade-offs remain useful once coding work becomes longer and less contained.
