How does HydraFusion decide which model to use?

HydraFusion uses capability signals for reasoning, code generation, debugging, and tool use to select the most efficient execution pattern to meet the quality bar. This can be a single model, a cascade, or a critique pattern.

What benchmarks were used to evaluate HydraFusion?

HydraFusion was evaluated on three agentic coding benchmarks: TerminalBench 2.1, DeepSWE, and CheckpointBench (an internal benchmark based on real GitHub Copilot sessions). Results varied by benchmark, with cost savings up to 67% and quality close to or better than Claude Opus 5.

AI Coding AssistantsLarge Language ModelsAI Business & IndustryGitHub Copilot Blog3h agoSep 5, 2026, 04:00 JST2 min read

GitHub unveils HydraFusion, a multi-model AI orchestrator

Send on LINE

Key takeaway

GitHub has launched HydraFusion, a research preview that automatically picks the best AI model workflow for each coding task.
The system can use one model, cascade to a stronger one, or have a critic review the output.
It aims to deliver high quality while reducing cost, showing big savings in tests.

3 Key Points

What happened
GitHub introduced Project HydraFusion, a research preview that automatically chooses and orchestrates models from multiple providers to handle coding tasks. It can draft, critique, and revise, or escalate to stronger models as needed.
Why it matters
In tests, HydraFusion delivered frontier-level quality while cutting costs: on TerminalBench 2.1 it improved verified task quality by 4.9 percentage points at 67% lower estimated cost compared to Claude Opus 5. On DeepSWE it came within 1.5 points of Opus 5 while reducing cost by 36%.
What to watch
HydraFusion is now in research preview for GitHub Copilot, designed for first-turn, single-prompt tasks. Developers can try it in autopilot mode and share feedback via /feedback in Copilot CLI or the GitHub Community discussion.

Ask the AI about this article →

Summaries like this, in your inbox every morning.

Get news like this every morning — free

Context & Analysis

HydraFusion represents a new approach to AI coding assistance: instead of just choosing a single model, it dynamically orchestrates multiple models from different providers to handle tasks. The system treats workflow selection as an optimization problem, balancing quality, cost, and latency for each request. While the examples here are technical, the underlying idea is about getting the best result without always paying for the most powerful model.

The research preview is part of GitHub's effort to bring automated semantic routing between local, cloud, and compound models to developers. Early internal testing suggests the capability is at or better than the comparison model, but the official results are mixed. On TerminalBench 2.1, HydraFusion improved quality while cutting cost, but on DeepSWE and CheckpointBench it was slightly lower in quality, though still at similar levels while reducing costs.

The preview is designed to learn which tasks benefit from compound workflows. For now, GitHub recommends starting with well-scoped, single-prompt coding tasks. The company plans to improve multi-turn performance next. HydraFusion remains an active research project, and its results, models, and availability may change based on feedback.

FAQ

How does HydraFusion decide which model to use?: HydraFusion uses capability signals for reasoning, code generation, debugging, and tool use to select the most efficient execution pattern to meet the quality bar. This can be a single model, a cascade, or a critique pattern.
What benchmarks were used to evaluate HydraFusion?: HydraFusion was evaluated on three agentic coding benchmarks: TerminalBench 2.1, DeepSWE, and CheckpointBench (an internal benchmark based on real GitHub Copilot sessions). Results varied by benchmark, with cost savings up to 67% and quality close to or better than Claude Opus 5.

GitHub Copilot BlogRead Original Article

Send on LINE

Also reported by GitHub Blog (AI)