Choosing an inexpensive model for a simple task sounds easy. Keeping that choice appropriate as a conversation becomes more demanding is harder. Copilot Fusion became generally available on October 28, combining automatic selection with Project HydraFusion's orchestration.

GitHub also announced an experimental preview of routing between local and cloud models. The distinction matters: Fusion is generally available, while automatic on-device routing remains experimental.

Effort becomes the main choice

Fusion appears in the model picker but represents an execution system. Standard targets everyday work; High and Max allow more intensive orchestration. The current documentation describes single-model execution, escalation after judging a draft, and review by other models before a final response.

This builds on HydraFusion, which we covered earlier. The product change brings those decisions into a shared experience across VS Code, CLI, the Copilot app and JetBrains, subject to subscription access and model policies.

For teams watching usage, the question changes. Rather than always picking the strongest or smallest model, they decide how much analysis a task deserves and inspect the outcome. Editing text and investigating a concurrency bug may warrant different levels of effort within the same project.

Local inference does not mean an offline session

The preview lets Fusion Standard consider compatible local models. GitHub describes managed installation through Windows ML and manual CLI configuration on other operating systems, with client-specific differences. Adding a model through a bring-your-own-key provider configuration does not automatically make it eligible.

Routing decisions still happen in the cloud. That rules out treating the feature as an offline solution or assuming the entire session stays on the device. Restricted code requires checking data paths and applicable policies first.

Hardware also affects the decision. A model appearing in the picker does not guarantee a good experience on every computer. Representative tasks should be assessed for latency, quality and resource consumption.

Savings depend on completed work

Billing follows the models used, so extra review calls can increase consumption. Documentation describes a temporary 30% model-cost discount for paid plans through November 30, 2026. A promotion can help an initial comparison without supporting a permanent cost forecast.

GitHub reports favorable benchmark comparisons for performance and cost. Without reproducing their conditions, those results should not become a promise of savings for every repository. A useful evaluation includes correctness, rework and review time alongside tokens.

One operational caveat deserves attention: documentation says discarded drafts do not automatically undo file edits already made in the workspace. A consolidated final answer can still leave files needing inspection. Model selection has become simpler; checking the diff remains part of the job.

Sources: Fusion announcement, local models and documentation.