GitHub Copilot Fusion Chooses When Coding Needs Another AI

A coding assistant does not need the same kind of help for every request. One task may be straightforward enough for a single model. Another may justify a more capable second attempt. A third may benefit from several models looking for mistakes before the answer comes back.

GitHub’s Copilot Fusion, released on October 28, puts those choices inside the assistant. Selecting Fusion means choosing a system that allocates work among AI models, rather than picking one model and leaving it in charge of everything. The interesting change is not simply which model wins. It is who gets called, for what purpose, and when.

One Model, an Escalation, or Several Reviewers

GitHub describes three ways Fusion can handle a request. In Single, one model does the work. In Cascade, an efficient model makes the first attempt and a judge evaluates the draft. The system can accept it or escalate the request to a stronger model.

Critique changes the relationship. A model drafts a response, several other models review it independently in parallel, and their combined feedback goes back to the original model for revision. They are reviewers, not merely replacements.

That distinction matters. An escalation asks a more capable model to take over. A critique asks additional models to challenge an existing answer. Both can add work, but they address different reasons a first attempt might be inadequate.

According to GitHub’s documentation, Fusion can choose a different pattern for the next prompt in the same conversation. Its choices depend on the task, the selected effort level and the available models. The documentation describes the process; it does not expose a complete decision rule that would let readers verify whether each individual routing choice was the best one.

The Extra Help Still Has a Cost

Fusion Standard selects one model for everyday work. High and Max can spend more model calls on demanding requests. Those extra reviews take time and can consume more AI credits. Asking several models is not free simply because the interface returns one answer.

GitHub charges for the models Fusion actually uses, at their respective rates, with no separate Fusion fee. Its documentation currently offers paid-plan users a 30% discount on model costs through November 30, 2026. That temporary offer is separate from any saving a routing strategy might achieve.

There are also reasons not to restart the whole conversation with every helper. GitHub says Fusion keeps the main conversation on the same model where possible to reuse cached context. A reviewer receives the context it needs, rather than necessarily receiving everything. Administrators can restrict the models available to Fusion; the system does not override the organization’s model policies.

A Promising Test, Not a Promise About Your Bill

Artificial Analysis tested Fusion’s Max setting in Copilot CLI. Its October 28 evaluation placed the system two points below Claude Code running Claude Opus 5.5 at maximum effort on the Coding Agent Index. Average model cost was $5.59 per benchmark task, compared with $11.42 for that Claude Code setup: about half.

These are pay-per-token API costs, charges for the text the models process and generate, not subscription prices, developer wages or the full cost of getting a change into production. The comparison is for the tested systems and settings, not a forecast of a particular team’s savings. It also measures the combined system, not how much each execution pattern contributed.

The results varied by task. Fusion led the test’s terminal-work component but did not lead its software-change component. A blended score is useful orientation, not a substitute for knowing which work needs to go well. As our earlier explanation of coding-agent costs showed, the software around a model and the way a test is run can change the outcome. Fusion adds another moving part: the choice of models within the workflow.

Rejecting an Answer Does Not Undo Its Changes

The most concrete limitation is easy to miss. GitHub warns that if Fusion discards a draft, file edits that draft already made are not automatically undone. Rejecting an answer and reversing its actions are different operations. Developers still need to inspect changes before committing them.

Nor does independent critique mean infallible critique. Separate model calls can still miss the same problem. GitHub also distinguishes Fusion from subagents: Fusion does not itself launch a team of separate task agents.

Fusion is therefore a shift from choosing a model to choosing a procedure for using models. Its value depends on whether the extra call is the right call: a stronger attempt when the first one falls short, a review when another perspective helps, and no extra machinery when one model is enough.