GitHub announced on November 17, 2026, that developers can select Copilot Fusion in the model picker in the latest Visual Studio versions. Fusion is a runtime system that chooses a model and execution pattern for each prompt, rather than asking you to pick a model every time. GitHub’s announcement covers the Visual Studio availability; its documentation explains how the routing works.

Treat effort as a budget choice: Standard uses one model per task, while High and Max can involve additional model calls and credits.

How Fusion routes each prompt

GitHub’s walkthrough shows Fusion’s general interface, effort levels and execution patterns

Fusion evaluates a prompt’s needs, including reasoning, code generation, debugging and tool use, then selects an execution pattern and eligible model. It also considers system health and model availability. Different prompts in the same session can take different paths, so a straightforward request need not follow the same workflow as a more substantial coding task.

PatternHow models work togetherHow the task proceeds
SingleOne modelA single model completes the task.
CascadeA drafting model and a judgeThe judge accepts the draft or escalates it to a stronger model.
CritiqueA drafting model and multiple reviewersReviewers assess the draft independently in parallel; the original model revises it using their combined feedback.

These patterns describe how Fusion can handle a prompt, not a fixed sequence for every request. GitHub’s walkthrough explains the effort levels and demonstrates Fusion’s general interface and execution patterns.

Effort levels, model additions and billing

Standard is intended for everyday work and selects one model per task. High supports more substantial coding tasks and can allow additional review calls; Max targets deeper reasoning. GitHub added MAI-Code-2 to Standard, GPT-6.1 Terra to High and Claude Sonnet 5.5 to Max. The models available to a user depend on their Copilot plan and applicable administrator or policy restrictions.

Fusion has no separate fee: GitHub bills the models it selects at their standard rates. High and Max can use additional calls, which may consume more AI credits than Standard. GitHub says paid Copilot plans receive a 30% discount on Fusion model costs through November 30, 2026.

What GitHub’s Terminal-Bench result measures

GitHub reports that Fusion Max achieved a 71.2% resolution rate on Terminal-Bench 4.1, a benchmark of realistic terminal tasks. The company reports a 5.4 percentage-point lead over Opus 5.5 and Fable 5.2, with benchmark costs 57% lower than Opus 5.5 and 88% lower than Fable 5.2. GitHub’s announcement reports these benchmark comparisons; the result measures terminal-task performance, not productivity in Visual Studio or performance on an individual team’s codebase.

Before committing changes, review the workspace: GitHub warns that discarding an intermediate draft does not automatically undo edits it has already made. GitHub’s Fusion documentation describes this behavior.

Local-model routing is a separate preview

Dark Copilot settings screen showing model-provider options and MAI Code 1.1 Flash downloading.

GitHub’s experimental local-model-routing preview names Copilot CLI, the Copilot app and Visual Studio Code. Even when a local model handles inference, Fusion’s routing decisions remain in the cloud, so the session is not offline. GitHub’s local-model announcement describes that preview and its client scope.

The Copilot model-provider settings shown in this interface include Windows ML, Anthropic, Azure OpenAI, Foundry Local and a custom endpoint. The screen also shows MAI Code 1.1 Flash downloading.