GitHub announces launch of Copilot Fusion at GitHub Universe 2026
November 4, 2026 · 2 min read
Also published in Русский, العربية, हिन्दी, Indonesia, Deutsch, Español, Français, Svenska, Português (Brasil)
GitHub announced the official launch of "GitHub Copilot Fusion" during its annual event, "GitHub Universe 2026," held in San Francisco on October 28 and 29, 2026.
Copilot Fusion is a multi-model feature that combines one or more AI models based on the characteristics of a given task. It evaluates the necessary capabilities for tasks such as inference, code generation, debugging, and tool usage, determining how AI models should be combined for optimal performance.
This new functionality eases the burden on developers, allowing them to manage the selection of AI models more efficiently while advancing development with a balance between performance and cost.
Copilot Fusion can select from one of three patterns based on the given task. The first, "Cascade," involves an AI model selected for efficiency creating a proposal for a solution, which is then validated by a quality assurance AI model that determines whether to escalate it to a more capable AI model.
The second pattern, "Critique," starts with one AI model drafting a solution, followed by an independent review by another AI model from a different family. If the review provides feedback, the original AI model revises the draft once based on this input.
To adjust the balance of quality and cost when selecting these patterns, different load levels are provided. "Fusion Standard" balances cost, quality, and waiting time, suitable for routine tasks. "Fusion High" offers balanced orchestration with moderate token savings, designed for well-defined, large-scale coding tasks that involve complex bug fixes or changes across multiple files. "Fusion Max" aims for cutting-edge performance while considering token efficiency, ideal for tasks requiring deep inference and thorough analysis. Lastly, "Fusion Fast" is selected for scenarios needing a high-speed model, which users cannot specify.
Additionally, there is a public preview option that allows executing tasks using locally installed MAI Code 1.1 Flash models or Qwen3.8 Flash Next models in Fusion Standard.