GitHub Copilot HydraFusion Cuts Cost 67% by Mixing AI Models
GitHub is moving Copilot from picking one model per request to orchestrating several models within a single task, and its own benchmark shows the mix cuts cost while lifting the completion rate.
Reporting from 2 sources: GameBusiness.jp, GIGAZINE.
GitHub announced Project HydraFusion on September 4 as a research preview for GitHub Copilot. The system routes a single task across multiple AI models, choosing between answering with one model, handing off to a stronger model, or having a second model review the output. In GitHub's TerminalBench 2.1 evaluation, HydraFusion cut estimated cost by 67% versus Claude Opus 5 while raising the correctly completed task rate by 4.9 points. It is available as an experimental feature in GitHub Copilot CLI on all plans.
Copilot already had automatic model selection that chose a suitable model per request. HydraFusion extends that by combining models within a single task. Depending on the request, the system answers with one model, hands the task to a higher-performance model when the first result is not enough, or has a second model review and correct the answer before it is returned. Developers only select HydraFusion; the routing is automatic. Billing is based on the actual token usage of each model called, at standard rates. GitHub plans to test the system on tasks that require long exchanges and says it may adjust the model lineup and routing based on what the research preview shows.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.