Copilot Fusion automates model choice

GitHub made Copilot Fusion generally available on 28 October. It chooses models and their workflow for each coding request, with Standard, High and Max effort levels. Supported clients include VS Code, CLI, the Copilot app and JetBrains.[1]

One model can answer directly; a draft can be checked before escalation; or reviewers can critique an answer before revision. Billing follows the models used, while subscription and organisation policies restrict the eligible pool.[1][2]

Discarding a draft does not automatically undo its file edits, GitHub’s documentation says. Reviewing accumulated code changes remains a separate step from accepting the final answer.[2]

Cost and capability differ by task

Artificial Analysis tested Fusion Max in Copilot’s command-line client. Its index score was two points behind Claude Code with Opus 5.5 Max; average API cost per task was $5.59 against $11.42.[3]

The composite weights three benchmarks equally. Fusion scored 71 on Terminal-Bench 4.0, leading the comparison, but 51 on FrontierCode, behind Claude Code’s 56. Those differences make the task mix more informative than a single headline ranking.[3][4]

The cost estimate uses token prices and caching charges. It excludes engineering, supervision and infrastructure costs, so it describes benchmark inference spending rather than a team’s total software bill.[4]

Report an error or suggest a correction