GitHub has introduced Project HydraFusion, a research preview that redefines how model choice is handled in coding assistance. Instead of relying on a single model for every request, HydraFusion dynamically constructs an execution plan tailored to each task. This approach allows the system to draft with one model, have another critique the output, or escalate to a more powerful model when initial results don’t meet quality standards. The models involved can come from different providers, and developers select HydraFusion just as they would any other model.
Availability and Deployment
HydraFusion is currently available as a research preview for users on all GitHub Copilot plans, but only within the GitHub Copilot CLI. There are no open weights or self-hosted deployment options. To enable it, users need to run /update, then /experimental on, followed by /model and select HydraFusion (Research Preview). Billing is based on the tokens consumed by the models invoked during each workflow, at each model’s standard rate.
How HydraFusion Works
HydraFusion builds on GitHub’s Auto model selection, which was released earlier in 2026 to match tasks with the most suitable model. HydraFusion extends this concept by treating workflow selection as an optimization problem. It analyzes capability signals related to reasoning, code generation, debugging, and tool use, then selects the least complex workflow that is expected to meet the quality threshold. This approach ensures that additional model calls are only made when they are likely to improve the outcome.
Three Execution Patterns
For each request, HydraFusion currently chooses one of three patterns:
- Single: A single selected model handles the task directly, preserving speed.
- Cascade: An efficient model drafts a solution, and a quality gate either accepts it or escalates to a stronger model if needed.
- Critique: One model drafts, an independent read-only critic from a different model family reviews the draft, and the drafting model revises once. This review process follows the same pattern as Rubber Duck.
Each pattern balances quality and cost differently. Single focuses on speed, Cascade maintains a path to stronger inference, and Critique adds an external perspective where review can outperform another unaided attempt.
Engineering Guardrails
GitHub has built the runtime around five key operating principles for repository-level work:
- Complete accounting: Every leg of the workflow—drafting, critique, revision, escalation, retry, and fallback—is fully tracked.
- Bounded execution: Each leg has explicit timeouts and cancellation options.
- Isolated review: Critics run in tool-less contexts and cannot modify the repository.
- Fail-safe application: No patch is applied if a workflow is cancelled or fails validation.
- Validated routing: Model bindings, fallback behavior, and availability are verified before execution begins.
Internally, the runtime logs role, outcome, cost, latency, and diagnostics for each leg. Externally, developers see a single coherent response and one permission-aware change set.
Benchmark Results
GitHub evaluated fixed HydraFusion policies on three agentic coding benchmarks, using Claude Opus 5 and GPT-5.6 Sol as baselines. All models ran at medium reasoning level. The reported figures are relative to Opus 5:
- TerminalBench 2.1: Estimated cost 67% lower, verified task quality +4.9 points.
- DeepSWE: Estimated cost 36% lower, verified task quality −1.5 points.
- CheckpointBench: Estimated cost 65% lower, verified task quality −0.1 points.
CheckpointBench is GitHub’s internal multi-turn benchmark, curated from real Copilot sessions and anchored to immutable public commits for replayability.
Key Takeaways
- HydraFusion selects a workflow per request, not just a model, across multiple providers.
- Three patterns are available today: Single, Cascade with a quality gate, and Critique with a cross-family reviewer.
- The best result was +4.9 quality points at 67% lower estimated cost on TerminalBench 2.1.
- On DeepSWE and CheckpointBench, it slightly trails Opus 5 while cutting costs by 36% and 65%.
- Available now in Copilot CLI via
/experimental, billed at each underlying model’s standard rate.
For more details, check the GitHub Blog announcement and the GitHub Community discussion #206492.
