GitHub Copilot Fusion launches, picking the best AI model per task

Three orchestration patterns

Fusion picks one of three patterns for every task. In one, an efficient model drafts a solution, a second model checks its quality and decides whether the job should be escalated to a more capable model. In another, one model writes a draft and a model from a different family reviews it independently. If the reviewer raises issues, the original model revises its solution exactly once.

To steer how quality and cost are traded off when a pattern is chosen, Fusion offers load levels. One is a balanced setting with moderate token savings, aimed at large but well-scoped coding work such as fixing complex bugs or changes that span several files. Another targets state-of-the-art performance while still watching token efficiency, and is meant for tasks that demand the deepest reasoning and the most thorough analysis.

Local models in preview

As a public preview, Fusion Standard can also run tasks on locally installed models. Developers can choose either MAI Code 1.1 Flash or Qwen3.8 Flash Next.

For teams already using Copilot, the change matters because model choice is increasingly a cost question as well as a quality one. By delegating that decision to the tool, GitHub is betting that developers will prefer a managed mix of models over picking a single one for every job.