GitHub Copilot Fusion can route a coding job across models instead of making you pick one

GitHub Universe’s new Fusion routing system claims near-frontier coding results at lower cost and with less manual model selection. It deserves a controlled benchmark against your current Cursor workflow.

At GitHub Universe on October 28, GitHub introduced Copilot Fusion, a generally available orchestration option that selects and combines models based on a coding task. Rather than asking a developer to decide whether a bug needs one model and a refactor another, Fusion offers Standard, High and Max effort settings. GitHub says its Max configuration approached Claude Opus 5.5 Max on a coding-agent benchmark while consuming roughly half the cost and taking about 40% less time. Those are vendor-reported comparisons, not independent results for your repositories.

For a solo developer shipping several products, the attractive possibility is not a benchmark trophy. It is fewer model-selection decisions and lower cost per completed task. The right metric is an accepted change that passes tests and needs little supervision. Your Sonos controller, Rinder, Wake Up and Stream Deck plugin are different kinds of work; a router might choose differently for TypeScript UI, server-side OAuth, and hardware integration.

There is an important Mac-specific nuance. GitHub also announced experimental routing between cloud and locally hosted models. On macOS and Linux, the supported route currently runs through Copilot CLI with a manually configured OpenAI-compatible local server; the managed model-download experience is Windows-only. The desktop app can select compatible local models directly, but its automatic local routing is presently Windows-focused. Fusion still contacts the cloud for routing decisions, so local inference does not mean offline operation.

GitHub says Fusion is rolling out to Copilot users and has a 30% usage discount for its first 30 days. A comparison must account for the user’s actual Copilot plan, credit rules and whether the option appears in the relevant interface. Avoid moving any production work to a new agent merely because a headline says it is cheaper.

A clean benchmark takes one real issue that can be reproduced automatically—perhaps a UI grouping edge case in HomeControl. Clone the branch twice, use the same written acceptance criteria, and run each agent from the same baseline. Compare the final diff, test pass rate, cost and minutes of human intervention. If Fusion wins, give it the kind of repeatable jobs where that advantage is real.

Run the same small, test-covered bug fix through Copilot Fusion and Cursor; record cost, elapsed time, correctness and how much intervention each required.