GitHub Copilot Fusion: how it picks models and what it costs

GitHub Copilot Fusion is an entry in Copilot's model picker that is not a model. Pick it, choose an effort level, and for each prompt Copilot decides which models to call and how: one model on its own, a cheap model whose draft gets checked and escalated if needed, or one model drafting while others critique. It has been generally available since 28 October 2026 in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs, It brings together two earlier Copilot features: auto model selection and the multi-model orchestration of Project HydraFusion.

The idea is Model RoutingModel RoutingSending each request to the cheapest model that handles it well, by rules, a classifier or a cascade, and checking with an eval that quality holds. done for you, per prompt, inside a coding agentAgentA model that uses tools in a loop, deciding each next step from what the last one returned, until the task is done.. Whether that beats choosing a model yourself depends on what you pay for and how much you want to know which model wrote your code.

How Copilot Fusion decides

For each prompt, Fusion looks at what the task needs, using signals for reasoning, code generation, debugging and tool use, and picks one of 3 patterns. GitHub says this choice adds minimal latency.

The pattern is picked per prompt, so a session can answer your first question with Single and the follow-up with Critique. Fusion also takes the models' current health and availability into account, and it switches models only at natural cache boundaries or when the conversation gets much harder, so a switch does not add cache costs without a better answer.

You see the progress of each step and the model running it, but only the final answer, since drafts can be revised or thrown away. To see which models answered, hover over the effort level name under the response.

Fusion Standard vs High vs Max

The effort level is the one thing you still choose. It sets how far Fusion will go, in model strength and in extra calls, to meet the quality bar.

• Fusion Standard balances cost, quality and speed and uses one model for each task. This is everyday work.
• Fusion High is for substantial, well-scoped coding jobs, such as fixing a complex bug or a change across several files, where extra model calls may be worth the time and credits.
• Fusion Max aims at frontier capability while still watching token use, for tasks that need the deepest reasoning.
• Fusion Fast cannot be picked. It runs only in VS Code's inline chat, the one place Copilot needs a fast model.

Not every plan gets every level:

If you were in the preview, nothing to redo: a saved choice of auto, HydraFusion or HydraFusion Max moves to the matching Fusion level.

What Copilot Fusion costs

Fusion has no price of its own. Each model it calls is billed at that model's standard rate, in the AI credits your plan uses. Two things follow.

On Standard, where Fusion picks one model per task, you pay for that one model. On High and Max, a task that goes through Cascade or Critique pays for every model in the chain, so the same task can use more credits than on Standard. To keep the main conversation cheap, Fusion keeps it on one model where it can so cached tokensPrompt CachingLetting the API keep the processed start of a prompt for a few minutes, so later requests that begin the same way are cheaper and quicker. keep counting, and reviewer models receive only the context they need.

Through 30 November 2026, paid Copilot plans get 30% off model costs when they use Fusion.

GitHub's headline number is from the Artificial Analysis Coding Agent Index: Fusion Max scored 60, within 2 points of Claude Code with Opus 5.5 Max, at about half the average API cost per task, about half the tokens and about 40% less time. That is GitHub reporting its own product. Run it on your own repository for a week and read your usage before you take it as your number. GitHub lists the rate of each model on its Models and pricing page for Copilot; for what the same models cost on their makers' own APIs, see our price tables.

When to pick a model yourself

Fusion takes a decision away, which is the point and also the cost. Pick a model by hand when:

• You need to know which model wrote the code. Fusion's pool changes as GitHub adds and evaluates models, and the patterns can change too. A hover shows the models after the fact, but you cannot pin one.
• The context is large. Fusion has no context window of its own. Each step runs inside the limits of its model, and the window Copilot shows for Fusion is the smallest among the models it can use. A long file set that fits one model's Context WindowContext WindowEverything a model can see at once while it writes an answer, counted in tokens. When the window is full, something has to be left out. may not fit Fusion's.
• The agent edits files. If Fusion throws away a draft, the edits that draft already made in your workspace are not undone. Review the changes before you commit, as you would with any agent, but especially on Cascade and Critique.
• You rely on subagentsSubagentAn agent started by another agent for one part of a task. It works in its own fresh context and returns only its findings.. Fusion works separately from subagents and does not start them.

For choosing a model by the job rather than by feel, our which model tool sorts current models by capability score and price. And if you are weighing Copilot against other coding agents, the Claude Code vs Cursor vs Codex comparison sets them side by side; Claude Code's own model settings are on choosing a model in Claude Code.

Local models on Fusion Standard

Fusion Standard can also send work to a model on your own machine, in public preview. On Windows, with a computer that meets the requirements, you can install MAI Code 1.1 Flash from Copilot CLI, the Copilot app or VS Code. In Copilot CLI on Windows, macOS or Linux, Qwen3.8 Flash Next installed through a provider such as Ollama works too, if Copilot sees it as qwen3.8-flash-next-local. Neither is used on High or Max.

Both need experimental features on: Local model auto routing in the Copilot app's Experimental settings, or --experimental (or /experimental on) in Copilot CLI.

Settings for admins

Fusion only calls models that your enterprise and organization model policies allow. It also skips models outside your plan and models excluded by data-residency, FedRAMP or evaluation-model policies. If none of its models are available to you, Fusion does not appear in the picker. Enterprise admins can see the models Fusion picks from and take one out of Fusion's pool without turning it off for everyone.

To make Fusion the default, set two keys in enterprise managed settings. A plain value locks the effort level against user and repository overrides; overridable sets a default that enterprise teams can change:

{
  "model": "fusion",
  "fusionTier": { "overridable": "standard" }
}

fusionTier takes "standard", "high", "max" or "unmanaged", and needs Copilot CLI 1.0.104-3, VS Code 1.144.0 or GitHub Copilot app 1.1.33-4 or later. Users can still pick another model for a single conversation.

A sensible start for a team is Standard as the overridable default, so everyday work stays on one model per task, and High or Max chosen on purpose for the jobs that earn the extra calls.