On September 4, 2026, GitHub announced ' Project HydraFusion ' as a research preview for GitHub Copilot, which uses multiple AI models depending on the task. According to GitHub's evaluation, in TerminalBench 2.1, it reduced estimated costs by 67% compared to Claude Opus 5, while also increasing the percentage of tasks that were completed correctly by 4.9 percentage points.
Project HydraFusion: Frontier quality via multi-model orchestration - The GitHub Blog
In AI-powered programming, the difficulty of the requested tasks varies greatly, from simple code generation to complex modifications. While high-performance AI models excel at difficult tasks, they often come with the challenge of increased processing time and usage costs.
GitHub Copilot has long had an automatic model selection feature that chooses the most suitable AI model based on the request. HydraFusion takes automatic model selection a step further, and it is said to be a system that combines multiple AI models as needed to process even a single task.
HydraFusion uses different methods depending on the request, such as 'answering with a single model,' 'transferring to a high-performance model if the initial model doesn't produce sufficient results,' or 'having another model review the answer before making corrections.' Developers only need to select HydraFusion, and the processing method and model to use are determined automatically.
GitHub has evaluated HydraFusion using multiple coding benchmarks. In TerminalBench 2.1, it showed a 67% reduction in estimated cost and a 4.9 percentage point increase in the percentage of tasks completed correctly compared to Claude Opus 5.
HydraFusion is available as an experimental feature of the GitHub Copilot CLI with all GitHub Copilot plans, and pricing is calculated based on the token usage and standard rates for each AI model actually invoked. GitHub says it plans to further validate it on tasks requiring longer interactions and may change the model configuration and processing methods depending on the results obtained from the research preview.
Related Posts:
in AI, Web Service, Posted by log1d_ts