Copilot Fusion

What it is

Copilot Fusion is a model selection and orchestration feature built into Copilot, GitHub's AI coding tool. Until now, developers had to choose for themselves which of several AI models to use. Because models differ in performance and price, a wrong choice could raise costs or lower output quality. Fusion automatically picks the model that suits the task from popular models such as those from OpenAI, Anthropic, Kimi and DeepSeek. It became generally available at GitHub Universe 2026. The ability to select local models as well is in experimental preview.

How it works

According to The Electronic Times, Fusion has three modes. 'Single' automatically picks one model suited to the task. 'Cascade' chains different models for each stage, such as planning and code implementation. 'Critic' cross-checks the output with a different model. Developers can set a tier that prioritizes either high-quality processing or fast, low-cost processing, but they do not need to specify a model for each task. If they want to use a specific model, they can turn off automatic selection and choose it themselves. Models not permitted by enterprise administrators are excluded from selection.

Why it matters

GitHub said, "You don't need the most expensive model to fix a single line of text on a website." The more AI agents a company uses, the higher its costs, and Fusion tries to balance performance and cost by giving easy tasks to cheaper models and hard tasks to high-performance models. GitHub said that in its own benchmark tests, Fusion achieved higher task completion scores at lower cost than using a single frontier model. These figures are the provider's claims. It also reflects a shift in the developer's role, from talking to a single model toward dividing work among multiple agents and coordinating them.

Things to know

According to GitHub's official changelog, Fusion can choose between on-device local models and cloud models in Copilot CLI, the Copilot app and VS Code. This feature is an experimental preview that users must turn on themselves. Running inference on a local model does not mean it works offline. The choice of which model to use is still made in the cloud. Local models may have different safeguards from the cloud models Copilot provides. Models connected directly (BYOK) also do not automatically become candidates for Fusion's selection.

Example

A company provides Copilot to its in-house developers and blocks a particular company's models in the admin settings. When a developer asks, 'Fix the bug on the login screen and add tests,' Fusion picks a suitable model from among the permitted ones. In Cascade mode, one model makes the plan and another handles the code implementation. In Critic mode, yet another model reviews the result once more. The developer does not have to track which models were used and can focus on reviewing the output.

Where it came up on this site

• briefing GitHub launches 'Copilot Fusion', which automatically selects and combines models for each task; local model routing in experimental preview 2026-10-28

Related terms

Coding agent Multi-agent LLM gateway Local LLM Token-based pricing AI agent

References

• "Ending the worry of choosing AI models": why GitHub built Fusion (The Electronic Times)
• [On-site report] Making games with AI and talking with robots: how GitHub expanded the developer ecosystem (ZDNet Korea)
• GitHub product VP: "We will lead the era of cost-effective AI coding" (ZDNet Korea)
• GitHub Copilot Fusion now supports selection of local models
• GitHub Copilot Fusion now supports selection of local models (The GitHub Blog)
• GitHub Expands Infra With 4 Mn Cores, Debuts Copilot Fusion for Multi-Model Routing (Analytics India Magazine)