GitHub Copilot Fusion and the Case for Multi Model Routing in Engineering Teams

GitHub moved Copilot Fusion to general availability on October 28, 2026, and the underlying idea is a direct response to a problem every team using multiple coding models already has: knowing in advance which model a given task actually needs is hard, and defaulting to the most capable model for everything is expensive and slow. Fusion combines two things GitHub had been running separately, a research project called HydraFusion for multi model orchestration and Copilot's existing automatic model selection, into one generally available feature that sits in the model picker next to naming an individual model directly.

Three workflows, not one model

Fusion operates at three effort levels. Standard targets everyday coding work with versatile, lower cost models. High targets harder problems that need deeper reasoning. Max targets the most demanding tasks with GitHub's most capable models available under your organization's policy. Underneath each effort level, Fusion runs one of three workflows. Single simply completes the task with one chosen model, functioning like ordinary auto selection. Cascade has an efficient model draft the response first, then runs a quality check that either accepts the draft outright or escalates the task to a stronger model when the draft falls short. Critique runs a draft and review loop between two models, where one model writes the first attempt, a second model reviews it, and the first model revises based on that review. Cascade is the workflow doing most of the cost and time saving in GitHub's own numbers, since a meaningful share of everyday tasks can be handled correctly by the cheaper drafting model without ever needing escalation.

Administrators get real controls here, not just a toggle. Individual models can be disabled for Fusion specifically without disabling them platform wide, and an organization can set any Fusion effort level as the default through enterprise managed settings, which matters for a team trying to control spend without removing choice entirely for individual developers. Billing follows the models actually used, each billed at that model's standard rate, with a 30 percent discount on model costs available for a limited time on paid plans.

What the benchmark comparison actually says, and what it leaves out

GitHub's reported numbers put Fusion Max within two points of Claude Code running Opus 5.5 at its own maximum effort setting on the Artificial Analysis Coding Agent Index, while claiming roughly half the average API cost per task, roughly half the tokens consumed, and about 40 percent less time spent per task. Those are GitHub's own figures, and the company has published an AT&T customer testimonial alongside them without disclosing the underlying evaluation methodology for that specific claim. Treat the cost and time multiples the way you would treat any vendor reported benchmark: directionally informative, not independently reproduced, and worth validating against your own codebase and task mix before you plan a budget around it.

The more durable point is structural rather than numerical. A routing system that drafts with a cheap model and escalates only when a quality check flags trouble is a legitimately different cost curve than always invoking your most expensive model, and it is the same bet several agent frameworks have been making independently, including the per worker model choices teams already make in multi agent automation, just without GitHub's distribution built in. Fusion also only routes between models your organization's enterprise or organization model policy already permits, so a tightly locked down policy limits how much benefit Cascade or Critique can actually deliver, since there has to be a real efficiency gap between the available cheap and expensive models for escalation to pay off.

Availability and what research preview users should know

Fusion is available now across all Copilot plans inside Visual Studio Code, the standalone GitHub Copilot app, GitHub Copilot CLI, and JetBrains IDEs, with support for additional clients described as coming soon. Anyone who was already using auto selection or either HydraFusion research preview is migrated automatically to the matching Fusion effort level, so there is no manual reconfiguration required for existing research preview users.

effort: max
workflow selected per task: cascade (draft) -> quality check -> escalate if needed
billing: per model actually invoked, at that model's standard rate

When this applies to your stack

If your engineering team is already running a mixed model strategy by hand, picking a cheaper model for boilerplate and a stronger one for architecture heavy work, Fusion's Cascade workflow is automating a decision your senior engineers are probably already making informally, and it is worth piloting on a defined slice of your backlog before rolling it out broadly. If your organization's model policy only permits one tier of model, the routing benefit mostly disappears, since there is nothing cheaper to draft with. That policy work gets harder when the underlying model lineup itself keeps changing, and it is worth stress testing any routing setup against your own coding agent pricing comparison rather than GitHub's. Contra Collective helps engineering organizations design the model policy, cost controls, and evaluation harness that make a routing system like this actually pay off rather than quietly becoming an expensive default, as part of our AI integration and agent engineering work.

FAQ

Does Copilot Fusion let me pick the exact model for a task? Not directly. You choose an effort level, Standard, High, or Max, and Fusion decides which models and workflow handle the task within what your organization's model policy allows.

How is Fusion billed? You are billed for each model Fusion actually uses, at that model's standard rate, with a 30 percent discount on model costs for a limited time on paid plans.

Is GitHub's cost and time comparison to Claude Code independently verified? No. The cost, token, and time multiples are GitHub's own reported figures, accompanied by a customer testimonial, with no published independent methodology.

What happened to HydraFusion and auto model selection? Both were folded into Fusion. Research preview users who had HydraFusion, HydraFusion Max, or auto selection enabled were automatically migrated to the matching Fusion effort level.

Can administrators restrict which models Fusion uses? Yes. Administrators can disable individual models for Fusion specifically without disabling them for other uses, and can set a default Fusion effort level through enterprise managed settings.