GitHub introduces Copilot Fusion — nearly at Claude Code level for half the cost

Copilot Fusion combines Project HydraFusion for coordinating multiple models and Copilot Auto Model Selection for automatically choosing them. The user specifies the system and effort level, and it determines which models to connect to a specific request and how to distribute the work among them.

In the Artificial Analysis Coding Agent Index, Copilot Fusion Max scored 60 points — less than 2 points below the result of Claude Code with Opus 5.5 Max. At the same time, the system used approximately half as many tokens, completed tasks approximately 40% faster and cost approximately half the average API cost per task.

Fusion can solve a task with a single model, verify the answer and pass it to a more powerful model, or organize the initial response, verification and refinement. The user chooses the Fusion Standard, Fusion High or Fusion Max levels, while administrators manage the available models and default settings. Charges are applied for each model used at its standard rate; for paid plans, a 30% discount on model costs is temporarily available.