GitHub expands HydraFusion model router to VS Code

GitHub has brought its Project HydraFusion research preview, previously restricted to the Copilot CLI, to the model picker in Visual Studio Code (version 1.140 or later) and the GitHub Copilot app. Although it looks like a standard model option, HydraFusion acts as an intelligent orchestrator. It analyzes task signals like reasoning, debugging, and code generation to select the most economical workflow that still satisfies its quality standards.

The router coordinates three distinct execution patterns: Single, Cascade, and Critique. Single uses one model directly. Cascade starts with an efficient model's draft and escalates to a stronger model if a quality gate fails. Critique involves one model drafting, a read-only critic from a different model family reviewing it without tools, and the original model making a final revision. The system is available to Copilot Pro, Pro+, Business, and Enterprise subscribers. There is no flat subscription fee; instead, users are billed at the standard token rates of the underlying models invoked during the run.

In offline evaluations, HydraFusion demonstrated significant efficiency gains compared to Claude Opus 5 at a medium reasoning level. On the TerminalBench 2.1 benchmark, the orchestrator achieved up to a 67 percent cost reduction. Results were more mixed on DeepSWE, which evaluates multi-file repository-level fixes; there, HydraFusion lowered expenses but trailed Claude Opus 5 in overall quality.

For software developers, HydraFusion automates the tedious process of manually switching between cheap and expensive models for complex tasks. However, the current preview is optimized for single-prompt, well-scoped tasks rather than long, iterative conversations. Additionally, intermediate drafts remain hidden during execution, though GitHub has added granular progress indicators to show the current stage of the orchestration cycle.