GitHub Copilot Employs Multi-Model Teaming: HydraFusion Sees Largest Cost Reduction of 67%

Dynamic Beating AI Newsflash: GitHub has added a multi-model orchestration system called HydraFusion to Copilot. It first determines how a task should be performed and then invokes different models.

There are currently three approaches: simple tasks are delegated to a single model for direct completion; for complex tasks, a lower-cost model is initially used, and if the results are unsatisfactory, it upgrades to a more powerful model; in cases requiring review, one model performs the initial task, while another model from a different family specifically identifies errors for the first model to rectify.

GitHub has benchmarked HydraFusion against Claude Opus 5 on three programming agent benchmarks. It outperformed by 4.9 percentage points on TerminalBench 2.1, lagged by 1.5 percentage points on DeepSWE, and only trailed by 0.1 percentage point on CheckpointBench, remaining relatively on par; cost reductions were 67%, 36%, and 65%, respectively.

This approach is reminiscent of Sakana AI's Fugu, as both elevate the decision of "which model to choose" to "how to organize multiple models." The difference is that Fugu itself is a trained orchestrator model that learns how to invoke different agents, and can even recursively call itself; meanwhile, HydraFusion is currently confined to selecting from three fixed execution modes – Single, Cascade, and Critique – more akin to directly integrating multi-model scheduling into Copilot.

HydraFusion has been made available for research preview to all GitHub Copilot subscribers, but currently requires activation in the Copilot CLI experimental features. GitHub also points out that at this stage, it is best suited for single-round programming tasks that can be clearly defined in one go, with ongoing optimizations for multi-round, long tasks.