GitHub Copilot Fusion selects the AI model for each coding task

GitHub opened Copilot Fusion, a system that selects the AI models for each coding task, to all users. GitHub tells that its top level had a score of 60 on a coding agent index. That is two points less than Claude Code with Opus 5.5 Max, at about half the average cost for each task.

How it works

The user selects Copilot Fusion in the model picker. Then the user selects a level: Standard, High or Max. Fusion uses three workflows:

• Single: one model solves the task.
• Cascade: a model with low cost writes a draft. A quality check can send the task to another model.
• Critique: one model writes a draft, a second model checks it, and the first model corrects it.

The numbers

GitHub tells that Fusion used about half as many tokens (units of text that a model reads). The bill is for each model at its standard rate. For a short time, paid plans get a 30% discount.

What is not known

GitHub gave no information about the test and no end date for the discount. These are results from GitHub. Administrators can limit the models and make Fusion the default.

Insight: GitHub decides which model maker gets the money

Fusion bills each model it uses at the standard rate. Customers of GitHub Enterprise pay for the AI credits they use. The result is that the router sets the bill. GitHub reads as the new gate between users and model makers, and it selects the model that gets the money. For leaders, the set of models that Fusion can use is a budget control.