GitHub Copilot Fusion reaches general availability for model routing
GitHub announced general availability of Copilot Fusion on October 28. Rather than selecting one named model for a task, users select Fusion and an effort level. The system then routes work to models and can orchestrate more than one model for a prompt.
The three effort levels are Standard for everyday coding, High for more challenging work and deeper reasoning, and Max for demanding tasks. These labels describe routing priorities, not fixed accuracy guarantees. A higher effort level does not establish that a generated change is correct.
GitHub describes three execution patterns: one model answering directly, a cascade that can escalate a draft after a quality check, and a critique pattern involving a draft, review and revision. Fusion combines earlier HydraFusion orchestration with automatic model selection. Developers still need to test and review the work that results.
Administrators can limit Fusion to models allowed by enterprise and organization policies. They can set defaults through Enterprise managed settings and disable a model for Fusion without disabling it for all other use. General availability does not override an organization's model permissions.
Usage is billed for each model invoked at that model's standard rate. The announcement offers a limited-time 30% model-cost discount for users on paid Copilot plans, without establishing a permanent discount. Its benchmark cost, token and time comparisons describe one reported test, not guaranteed savings on every team's workload.
Fusion is available on all Copilot plans in Visual Studio Code, the Copilot app, Copilot CLI and JetBrains IDEs, with client updates needed. GitHub says other clients will follow. Existing research-preview selections transition to corresponding effort levels, so teams should check the current model picker and policies before assuming their previous selection is unchanged.