GitHub has made Copilot Fusion generally available, replacing the usual one-model-per-task choice with a routing system that can select and combine models during a coding session. The architectural change matters because quality, latency, token use, and the bill can now depend on an orchestrator rather than a single model selected by the developer.
What changed
Fusion offers Standard, High, and Max effort levels. For each prompt it can use one model directly, let an efficient model draft and escalate only when a quality check fails, or have one model critique another model’s answer before revision. Administrators can limit Fusion to models already permitted by enterprise or organization policy and can make an effort level the default.
Fusion is available on all Copilot plans in Visual Studio Code, the GitHub Copilot app, GitHub Copilot CLI, and JetBrains IDEs. Existing preview selections for auto or HydraFusion transition to corresponding Fusion levels.
The practical consequence
Developers no longer need to guess which model fits every step, and enterprises can centralize that tradeoff. The cost model also changes: GitHub bills each model Fusion invokes at that model’s standard rate. Paid plans receive a temporary 30% discount on Fusion model costs, but GitHub documentation says the discount runs through November 30, 2026, so teams should not treat it as a permanent budget assumption.
Limits and evidence
Routing is not a guarantee of a better answer, and multi-model critique or escalation can consume more than one model invocation. Client support is not universal yet, and administrators must review allowed-model policies if data handling or provider choice is constrained.
GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, while using about half the tokens and API cost and taking about 40% less time per task. GitHub reported those comparisons; DrComps has not independently reproduced them, and real project results can vary with repository context, tools, model policy, and effort level.