GitHub Copilot Fusion: Pricing, Model Routing and Local Model Limits

GitHub made Copilot Fusion generally available on October 28. It chooses models and coordinates their work instead of asking developers to select a model for every task. GitHub also introduced experimental local-model routing, but that does not turn Copilot into an offline coding assistant.

What Copilot Fusion changes

The launch announcement combines the earlier HydraFusion orchestration work with automatic model selection. Fusion is available across Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs, subject to the models permitted by the user's plan and administrator.

The controls are Standard, High and Max effort. The developer expresses how demanding the task is, while the system chooses how to perform it. That is a different interface from deciding that one particular model should handle an entire session.

For a team, the practical question becomes whether the router makes sensible decisions across its actual workload. A small rename and a difficult debugging task should not automatically receive the same amount of computation.

How the model routing works

GitHub's Fusion documentation describes three patterns: a single model completes the task; a cascade checks an efficient model's draft and may escalate; or a critique workflow gathers reviews before revising the answer.

Fusion has no independent context window; each step uses its model’s limits. Standard is available on every plan, while High and Max require Pro+, Max, Business or Enterprise.

One limitation is especially important: discarding a draft does not automatically undo file edits that draft already made. A better final answer does not guarantee a clean working tree. Developers still need to inspect the resulting diff before committing.

That makes repository state part of the evaluation. Testing only the final chat response could miss a change left behind by a discarded attempt.

Copilot Fusion pricing

Fusion charges for the models it uses rather than adding a separate Fusion fee. GitHub's documentation specifies a 30% discount on model costs for paid Copilot plans through November 30, 2026. High and Max can consume more credits than Standard because they may involve additional model calls.

A discount is not a fixed task price. A useful comparison is total spend for a successfully reviewed change, including any retries, rather than the cost of the first response alone.

Local model routing is an experimental preview

The separate local-model announcement describes opt-in routing between supported device models and cloud models through Fusion Standard.

Managed MAI Code 1.1 Flash downloads use Windows ML on supported Windows hardware. Copilot CLI can also route to manually configured supported models on Windows, macOS and Linux, including MAI Code 1.1 Flash and Qwen3.8 Flash Next through a local OpenAI-compatible server.

The app's intelligent local routing currently supports managed Windows ML models only. Directly selecting a compatible local model through BYOK is a different capability; discovering a model does not automatically make it eligible for routing.

Local inference does not mean offline operation. Routing decisions still happen in the cloud, and local models may have different safety protections. Teams with strict data-handling requirements should assess the entire request path, not infer privacy guarantees from where one inference step runs.

What to test before making Fusion the default

Start with representative tasks and compare the changed files, completion time and total credits consumed. Include work where an incorrect edit is easy to overlook, because an extra model review is useful only if it catches the mistakes that matter.

Fusion moves model selection into the coding system. The next competitive question is how well that system can spend less on easy work without making difficult work harder to trust.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn technical releases and developer workflows into clear launch stories and founder-led content.

Book a 30-minute call.