GitHub introduced Project HydraFusion as a research preview following Auto model selection, which enables developers to automatically select the model suited to their tasks. HydraFusion completes tasks by selecting between models from multiple providers at runtime, drafting, critiquing, revising, or handing them off to more capable models. The system automatically routes tasks among local, cloud, and compound models while balancing performance, cost, and latency.
HydraFusion is available in GitHub Copilot CLI, which users on all GitHub Copilot plans can access via /experimental. Users need to update with the /update command, enable the /experimental feature, and select HydraFusion (Research Preview) from the /model menu. Pricing is based on the standard token prices of the models used.
For each request, the system selects one of three operating modes: “Single,” in which a single model directly produces a solution; “Cascade,” in which the draft produced by an efficient model undergoes quality control and is handed off to a more capable model if necessary; and “Critique,” in which a model’s output is reviewed and revised by an independent critic from a different model family.
In offline evaluations across three coding benchmarks, HydraFusion demonstrated frontier-level quality at a low estimated cost. On TerminalBench 2.1, it increased verified task quality by 4.9 points compared with Claude Opus 5 while reducing the estimated cost by 67%.
Why it matters
This approach is changing the single-model-dependent way of working in coding processes by reducing the burden on developers of choosing a model for each task. Routing local, cloud, and composite models based on performance, cost, and latency allows users of Copilot CLI in particular to benefit from different quality and resource options within the same workflow. Although offline benchmark results indicate that high quality is being targeted at a lower estimated cost, these findings alone do not explain how they translate into the real-world use of the system, which is a research preview. The key open question for users is how this task-dependent routing will affect the total cost based on standard token prices and the day-to-day development experience.