Copilot Fusion launches with automatic model routing and review

GitHub launched Copilot Fusion on October 28. Developers choose an effort level, while Fusion selects the coding models that draft, review or revise a response.

Available in VS Code, the Copilot app, CLI and JetBrains IDEs, Standard targets everyday coding, High deeper reasoning, and Max demanding tasks. All plans are supported; preview selections transition to corresponding effort tiers.

The execution patterns are a direct answer, a checked draft accepted or escalated to a stronger model, and a draft revised after criticism.

Users can follow model progress and inspect the models used below a completed response. Only the final response is shown. Discarding a draft does not automatically undo file changes it already made in the workspace.

Fusion carries no separate charge. Each model used is billed at its standard rate, so High and Max may consume more AI credits than Standard. Paid Copilot plans receive a 30% discount on model costs through November 30, 2026, according to the current documentation.

The cost behind model selection

GitHub prices input, output and cached tokens separately, then converts consumption into AI credits. One credit equals $0.01. Individual plans include allowances; Business and Enterprise pool per-user allowances at the billing entity. Additional usage beyond included allowances follows published per-token rates.

For example, the current default-context rates for Claude Sonnet 5.5 are $2 per million input tokens, $0.10 for cached input and $10 for output, with cache writes separately priced at $2.50. Actual token quantities and cache use determine the charge; these are not flat prices for a completed Fusion task.

Enterprise administrators can configure Fusion’s model pool separately from general model availability. A model enabled for Fusion must also satisfy the user’s other model policies. Automatic routing therefore operates inside the organization’s permitted access, with a distinct configuration surface for the orchestrator.

An enterprise-wide baseline can be supplemented by delegated organization decisions or targeted organization rules. An opt-in preview moves granular access to enterprise teams. Switching to that mode deactivates organization-level model policies and requires delegated models to be configured for the relevant teams.

Manual model selection in CLI persists when a session resumes. A change applies immediately while idle or after the current turn, without affecting other sessions. Choosing Fusion instead delegates model and execution-pattern selection for each prompt. GitHub describes different model strengths in latency, reasoning and context handling. Project settings, enforced policies and explicit invocation options can override a default model. Setting a default model applies to the current and future sessions, while other existing sessions retain their selections.

A separate October 28 release put local-model routing into experimental preview. Supported Windows devices can download and configure MAI Code 1.1 Flash, then let Fusion Standard choose between on-device and cloud models. The Copilot app’s intelligent routing currently supports managed Windows ML models. That experimental scope applies to the on-device extension of Fusion.

In CLI, manually configured local routing also works on Windows, macOS and Linux with supported models. Users manage the download, inference runtime and device resources for MAI Code 1.1 Flash or Qwen3.8 Flash Next. Merely discovering a model through a provider configuration does not make it eligible. Routing decisions still happen in the cloud, so local inference does not make the session offline.