Project HydraFusion is GitHub’s research preview for runtime multi-model orchestration in GitHub Copilot. Instead of asking developers to decide which AI model should handle each coding task, HydraFusion evaluates the request and constructs an execution workflow using models from multiple providers.
The approach is designed around three goals: maintaining strong task quality, controlling workflow cost, and managing latency. In controlled offline evaluations, selected HydraFusion configurations reached or approached the quality of the evaluated Claude Opus 5 baseline while using substantially less estimated workflow cost.
Key points:
-HydraFusion dynamically selects models and execution patterns for each request.
– It can use direct solving, escalation, or independent critique and revision.
– Offline testing covered TerminalBench 2.1, DeepSWE, and CheckpointBench.
– HydraFusion is available as a research preview through GitHub Copilot CLI.
How HydraFusion works
HydraFusion treats model selection as an optimization problem. It considers capabilities such as reasoning, code generation, debugging, and tool use, then chooses the least complex workflow expected to satisfy the task.
There are currently three main execution patterns:
- Single: one selected model attempts the task directly when additional orchestration is unnecessary.
- Cascade: a more efficient model creates the first solution, after which a quality gate determines whether the result is sufficient or should be escalated to a stronger model.
- Critique: one model produces a draft, while an independent model from another family reviews it without modifying the repository. The drafting model then revises the result once.
This selective design matters because not every coding request needs multiple model calls. Simple tasks can stay fast, while difficult tasks can receive additional reasoning or review when the expected benefit justifies the added cost.
Building a reliable orchestration system
Combining several models into one coding experience requires more than routing logic. HydraFusion is built around controls that keep execution predictable and repository changes safe.
Its operating principles include:
- Complete accounting: costs and token usage are tracked across drafting, critique, revision, escalation, retries, and fallback steps.
- Bounded execution: every workflow leg has defined timeout and cancellation behavior.
- Isolated review: critic models work in separate, tool-free contexts so they can assess proposed work without changing the repository.
- Fail-safe application: incomplete or failed workflows do not apply unvalidated patches.
- Validated routing: workflow definitions, model bindings, fallbacks, and model availability are checked before execution starts.
These controls allow developers to see one coherent result even though multiple model calls may have happened behind the scenes. The runtime records workflow roles, outcomes, cost, latency, and diagnostics for analysis.
Benchmark results
GitHub evaluated fixed HydraFusion policies on three agentic coding benchmarks: TerminalBench 2.1, DeepSWE, and CheckpointBench. Claude Opus 5 and GPT-5.6 Sol were used as comparison baselines. The evaluations kept task inputs, tools, execution limits, pricing assumptions, grading conditions, and treatment of missing results consistent across configurations.
Relative to Opus 5, the best tuned HydraFusion configuration showed:
- TerminalBench 2.1: 67% lower estimated cost and 4.9 percentage points higher verified task quality.
- DeepSWE: 36% lower estimated cost and 1.5 percentage points lower verified task quality.
- CheckpointBench: 65% lower estimated cost and 0.1 percentage points lower verified task quality.
The results are controlled offline measurements rather than guarantees for every developer workload. They depend on the evaluated benchmark versions, model pool, workflow configurations, pricing assumptions, and reasoning settings. The research preview is intended to test how these findings translate into real-world usage.
What the benchmarks show
TerminalBench 2.1 focuses on complex, multi-step terminal tasks. It provided the clearest sequence for observing HydraFusion’s iterative development and showed a strong quality-cost result.
DeepSWE measures repository-level software engineering work involving large codebases, cross-file dependencies, and end-to-end fixes. HydraFusion remained close to Opus 5 while reducing estimated cost by more than one-third.
CheckpointBench is an internal multi-turn benchmark built from real GitHub Copilot agentic coding sessions. Sessions are tied to public repositories and immutable commits, making them replayable. The dataset covers different programming languages, task types, and difficulty levels, including debugging, feature development, explanations, configuration, planning, refactoring, and testing.
Together, the benchmarks are intended to measure whether orchestration can preserve strong coding performance without relying on the most expensive model for every step.
How HydraFusion was improved
The routing policies were refined through repeated experiments across CheckpointBench, DeepSWE, and TerminalBench 2.1. Rather than tuning the system around one benchmark, GitHub compared candidate policies across several evaluation sets.
HydraFusion uses capability scores to compare routing options. Beam search was then used to explore candidate decision policies and identify combinations that balanced quality, cost, and failure behavior against a frozen baseline.
The development process also exposed operational problems in the evaluation harness. Two invalid runs were excluded after failures were identified and corrected. After those issues were resolved, later configurations continued improving. This illustrates why reliable evaluation infrastructure is important when optimizing compound AI workflows.
Research preview and what developers can try
HydraFusion is available as a research preview to GitHub Copilot users through Copilot CLI’s experimental features. Developers can update the CLI, enable experimental functionality, and select HydraFusion as the model.
For the current preview, GitHub recommends starting with substantial, clearly defined coding tasks that can be given to Copilot in a single prompt and handled in autopilot mode. The team is particularly interested in learning how orchestration performs on longer, multi-turn sessions.
Developers can provide feedback through Copilot CLI or the GitHub Community discussion. Feedback will help shape progress visibility, workflow behavior, latency, reliability, caching, cost, and safety.
Benchmark comparison
| Benchmark | Estimated Cost vs. Opus 5 | Quality vs. Opus 5 |
| TerminalBench 2.1 | 67% lower | +4.9 points |
| DeepSWE | 36% lower | -1.5 points |
| CheckpointBench | 65% lower | -0.1 points |
Conclusion
Project HydraFusion represents a shift from simply selecting a powerful AI model to dynamically constructing a workflow for solving a coding task. Its central idea is that the best result may come from combining models strategically rather than using one frontier model for every request.
The research preview provides an opportunity to test that idea with real developer workloads. As new models become available, HydraFusion can evaluate them and add their capabilities to the orchestration pool. GitHub’s ongoing work will determine how effectively this approach can deliver strong coding quality while reducing unnecessary model usage, cost, and latency.
Read more such articles from our Newsletter here
