your123 - stock.adobe.com
GitHub tests multi-model routing with HydraFusion
GitHub Copilot's HydraFusion dynamically selects workflows to balance cost and quality.
GitHub's new multi-model router, HydraFusion, optimizes workflows based on developer input to balance between output quality and AI costs.
The tool creates an execution plan and chooses a model – or multiple models – based on the task with which a developer prompts GitHub's Copilot AI assistant.
HydraFusion – and the multi-model routing approach in general – are popular because they help to rein in agentic workflows amid rising AI costs for enterprises, Jason Andersen, VP and principal analyst at Moor Insights & Strategy told TechTarget.
"Since token costs are a major driver for developers right now, model routers help customers save significant money," Andersen said. "The cost variability is massive in some cases, depending on your input and the required outputs."
HydraFusion builds on Copilot's Auto model selection feature from earlier this year. The GitHub blog describes HydraFusion as its "first bet" on an approach to agentic development, "moving from choosing the best model to dynamically constructing the best way to solve each task." The Auto tool chose one model for each workflow, while HydraFusion chooses one or several based on the workflow type.
Usage is based on the tokens consumed by the models HydraFusion uses, priced at each model’s standard rate, according to the GitHub community page.
What sets HydraFusion apart?
Many vendors have their own versions of a model router. Cursor and IBM are two examples.
"Model routers have been around for a while now," Andersen said. "Hyperscalers like AWS and Azure have their own, but are only aligned to the models they support. OpenRouter -- which is being acquired by Stripe -- and Switchyard from NVIDIA open routers up across different clouds."
HydraFusion might have an advantage over these offerings because it offers platform independence that the others do not, according to Andersen. HydraFusion will integrate local and remote models. New frontier models will be incorporated in HydraFusion's model pool as they become available in GitHub Copilot.
"GitHub may offer the broadest choice of models and how they are hosted. For GitHub, this is really important, since its value proposition has always been choice and cross-platform support," Andersen said.
Allowing users to bring models into the orchestration layer could be valuable, said Ned Bellavance, founder and technical educator at Ned in the Cloud.
"People get really invested in a particular harness – whether it's Claude Code or VS Code or Antigravity," Bellavance said. "They're going to want to bring these other models into that harness. So it's interesting to see the ones that are embracing that 'pick any model you want' [approach]."
HydraFusion's execution patterns
HydraFusion scores a task and routes it to one or several models. HydraFusion builds on Hybrid Dynamic Routing Architecture (HyDRA), GitHub's existing routing framework, which scores tasks across reasoning, code generation, debugging and tool use. HydraFusion uses the scores HyDRA creates to pick an orchestration pattern.
The orchestration patterns HydraFusion uses to evaluate tasks also might give it a subtle edge. Depending on the evaluated task, the router can choose between three approaches.
In a "Single" approach, the router can select one model to see the task through.
In a "Cascade" approach, the router drafts and tests a solution with an efficient model. If the model fails the test, the task is sent to a more capable model.
In a "Critique" approach, one model develops a draft, then sends it to an independent reviewer model from a different family for a read-only critique. The reviewer cannot alter the draft, just comment on it. It sends the comments back to the first model, which guides that model's revisions. This approach is akin to rubber duck debugging.
"We are quickly learning that the economics of using LLMs for every coding task often do not work, and also that LLMs are hard to control for consistency of results," said Torsten Volk, an analyst at Omdia, a division of Informa TechTarget. "Cursor also does agent routing and workflows, but not quite to that degree where it selects workflows based on its evaluation of a task. It's really a subtle difference."
The authors of the blog announcing HydraFusion's release said that HydraFusion's routing processes are based on how developers use GitHub Copilot in the wild, and refined against several benchmarks including CheckpointBench, DeepSWE and TerminalBench 2.1.
HydraFusion is still an active research effort. Developers engaging with the research preview were encouraged to share debug logs or their session ID when providing feedback.
"The value of HydraFusion is hard to gauge without seeing how it does out in the wild. I would imagine that it's just a starting point for GitHub to show the value of Copilot as an orchestrator for multi-agent workflows," Volk said.
Ben Lutkevich is an award-winning writer and editor for TechTarget.com
