Tool Suggestions
GitHub Copilot Fusion: Stop Picking Models — It Routes the Task for You
GitHub made Copilot Fusion generally available on October 28, and it changes one decision developers make dozens of times a day: which model to use. With Fusion you pick an effort level — Standard, High or Max — and it routes each task across models itself, escalating when a cheap attempt fails and adding critique reviews between models.
How it works
Fusion evaluates what a task needs — reasoning, code generation, debugging, tool use — then selects one or more models to do it. A failed cheap attempt escalates to a stronger model; a second model can review the first’s work. It is available in VS Code, the GitHub Copilot app, the CLI and JetBrains IDEs. On Fusion Standard with experimental features, it can even choose models installed on your machine, including MAI Code 1.1 Flash on qualifying Windows computers and Qwen3.8 Flash Next in Copilot CLI.
The headline claim
GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max (62), at about half the average API cost per task ($5.59 against $11.42) and about 40% less time per task. Satya Nadella’s version goes further: frontier quality at up to 77% lower cost. Treat both as vendor claims until independent benchmarks confirm them — and note what is not stated in GitHub’s materials: whether GitHub commissioned or funded the evaluation. The mechanism itself is sound: most coding tasks do not need a frontier model, and the waste comes from humans defaulting to the expensive model “just in case”.
What else landed at Universe
Copilot code review Max, which GitHub says posted the highest F1 score on ReviewBench for high-severity issues (Lite and Balanced remain for routine reviews). Copilot Automations that learn from their outcomes and propose improvements for approval. And a billing change: from October 28, GitHub Enterprise customers can use all Copilot features on an AI-credit model instead of separate subscriptions. GitHub’s Octoverse report gives the context: agents started more than 40% of pull requests in September, up from 17% in March — and GitHub is adding 4 million CPU cores and 160 petabytes of storage to handle the agent surge.
What to watch before switching
High and Max can spend more than Standard, so set effort levels per repo and task type the way you set CI budgets, and measure your own cost per merged pull request rather than trusting vendor charts. A discarded draft will not roll its edits back. And read the Fusion docs before making it your organization’s default. The skill that matters now is not model-picking but task specification: intent, constraints and acceptance criteria, written clearly — a vague task routed brilliantly still produces vague code.