GitHub announced Copilot Fusion as generally available on October 28. Instead of choosing a single model for every prompt, users select an effort level and let Fusion choose and coordinate models for a task.
Three execution patterns
- Single sends the task to one model.
- Cascade starts with a more efficient model and can escalate when a quality check calls for it.
- Critique has one model draft an answer and another review it before a revision.
Administrators control which models are available to Fusion, and the announcement says billing follows the standard rate for each model used. The product is available through several Copilot clients, with feature availability tied to plan and client support.
The practical question is whether automatic routing improves outcomes enough to justify the cost and latency for a team's common tasks. Compare representative work and review the model-use details before making it a default.