GitHub brings model routing into Copilot Fusion

GitHub Copilot Fusion brings model selection into the coding workflow.

GitHub announced Copilot Fusion on October 28 at GitHub Universe. Fusion combines model routing with multi-model orchestration. Developers select a level of effort, Standard, High or Max, and Fusion handles model selection for the task. The announcement moves a decision developers often made in the model picker into Copilot's handling of the task itself.

A model picker gives a developer a visible choice at the start of a session. A router can choose based on what the developer is asking it to do. Summarizing a diff, tracing an unfamiliar call path and proposing a patch place different demands on a system. Routing gives teams a way to compare results across those ordinary tasks without having to assign one model to all of them.

Measuring the value of routing

GitHub cites AT&T as a customer that saw improved token efficiency with Fusion Standard. That account reflects the customer's own workload. A team running its own trial can compare similar work across manual selection and routing, then record accepted changes, follow-up prompts, time spent reviewing and usage. A session that takes fewer tokens but requires a reviewer to undo more changes can still cost more engineering time.

GitHub also announced a connected cloud experience for using Copilot across GitHub and mobile. The two changes raise different adoption questions. A routing evaluation focuses on output, cost and model policy. A cloud workflow evaluation focuses on repository access, sandbox usage and where a developer reviews the proposed changes.

What to verify in a pilot

GitHub lists Fusion as generally available on all Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs. GitHub applies standard usage rates for the models involved and announced a limited 30% discount for paid plans. Administrators should verify their effective model policies and the current billing terms before expanding a pilot. A small set of known repository tasks provides a baseline that a team can rerun. Inspect the accepted patch and its tests before comparing usage so that a faster session with a weaker result does not count as a productivity gain.

[GitHub Copilot Fusion changelog](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/) · [GitHub Universe announcement](https://github.blog/news-insights/company-news/github-universe-2026-evolving-every-layer-of-the-stack/)