GitHub shipped a research preview this month that quietly changes how Copilot CLI works under the hood. Project HydraFusion stops sending every coding task to a single model and instead builds a runtime execution plan, routing each request across whichever combination of models best fits the job. On TerminalBench 2.1, the strongest configuration cut estimated costs 67% against Claude Opus 5 while scoring 4.9 points higher on verified task quality. It is available today on all Copilot plans through a single CLI flag.
Three Patterns, One Router #
HydraFusion does not expose model selection to the developer. You select it the same way you would select any other model, and it decides internally which of three execution patterns to use for your specific task.
- Single: One model handles the request end to end. Used for straightforward tasks that do not require multi-step reasoning or revision.
- Cascade: A cheaper, faster model drafts a solution first. A quality gate then evaluates the output. If it clears the bar, the draft ships. If it fails, the request escalates to a more capable model to finish the job. The expensive model activates only on the subset of requests that fail — not on every call.
- Critique: One model drafts, then an independent critic from a different model family reviews it. The original drafter revises once based on that critique. This cross-provider review catches reasoning gaps that a same-family critic would miss.
The Cascade pattern is where the economics get interesting. Because the frontier model only fires when the cheap model fails the quality gate, most tasks never touch the expensive tier. That is the mechanism behind the cost reductions.
The Benchmark Numbers, Read Honestly #
GitHub published results across three benchmarks, all compared against Claude Opus 5 as the baseline:
| Benchmark | Quality vs Opus 5 | Cost vs Opus 5 |
|---|---|---|
| TerminalBench 2.1 | +4.9 pts | ‑67% |
| CheckpointBench | ‑0.1 pts | ‑65% |
| DeepSWE | ‑1.5 pts | ‑36% |
Costs are down across all three. Quality holds on two of three — TerminalBench improves, CheckpointBench is essentially even at 65% cheaper. The real exception is DeepSWE, where HydraFusion trails Opus 5 by 1.5 points. VentureBeat’s headline called this “costs down in every benchmark, quality matched in one,” which is technically accurate but buries the CheckpointBench result. The honest read: a strong trade on most workloads with one real gap in the DeepSWE category.
How to Enable GitHub HydraFusion #
HydraFusion is live in GitHub Copilot CLI experimental mode. To enable it:
- Update GitHub Copilot CLI to the latest version
- Run
/experimental onin the CLI - Run
/modeland select HydraFusion from the list
Available on all Copilot plans — Free, Pro, Business, and Enterprise. Usage billing is based on the tokens consumed by whichever models HydraFusion routes to, at each model’s standard rate. See the official GitHub announcement for full details.
Why This Differs From Building Your Own Routing Layer #
Multi-model orchestration is not a new idea. LangGraph is already running in production at companies like Klara and Replit. Bifrost offers enterprise-grade routing with sub-millisecond overhead. The difference with HydraFusion is that there is nothing to build or integrate — the routing layer is already inside your Copilot CLI. For teams that would otherwise spend engineering time constructing an orchestration layer, the zero-configuration approach is the actual value proposition, not just the cost numbers.
The broader pattern is worth noting. Over 2.4 billion API calls per week were routed through multi-model frameworks in enterprise environments in Q1 2026. The architecture is proven at scale. HydraFusion brings that same pattern to individual developer tooling — no infrastructure required.
What to Know Before Enabling It #
HydraFusion is a research preview. Some developers have reported errors when activating it via the experimental flag — the community discussion thread has open questions GitHub has not yet answered. It is currently CLI-only: no HydraFusion support in the IDE extensions yet, and no committed timeline for that. The DeepSWE quality gap is a real consideration if your work aligns with that benchmark category.
For teams already paying for Copilot, testing costs nothing. The routing logic runs transparently, quality gates are automatic, and the model pool spans multiple providers without any configuration on your end. If TerminalBench and CheckpointBench map closer to your actual work than DeepSWE does, the numbers argue for turning it on. A step-by-step setup guide is available if you hit errors with the experimental flag. The more meaningful signal is the direction of travel: GitHub is building the routing layer into the product, not leaving it as an infrastructure problem for teams to solve independently. That shifts the question from “which model should I pick” to “which orchestration strategy fits my workflow” — and that is a more interesting question to be asking in 2026.