GitHub Copilot Fusion now picks your model for you

Your model picker just got a default worth trying.

GitHub made Copilot Fusion generally available on Oct. 28. Instead of choosing a model, you choose an effort level, and Fusion decides which models handle each prompt and how they work together. It is on every Copilot plan.

You pick the effort, GitHub picks the models

GitHub offers three effort levels. Standard uses versatile models for everyday coding. High uses more capable models for harder problems. Max uses the most capable models for the most demanding tasks.

Under each level, Fusion runs one of three workflows, according to the changelog. In Single, one model does the task. In Cascade, an efficient model drafts and a quality check either accepts the draft or escalates to a stronger model. In Critique, one model drafts, a second reviews, and the first revises.

Fusion combines the multi-model orchestration from GitHub's Project HydraFusion research preview with Copilot's auto model selection. If you had turned on auto, HydraFusion or HydraFusion Max, GitHub says you move to the matching Fusion level automatically.

The same day, Microsoft introduced the Copilot harness, the orchestration layer GitHub's engineers built and Microsoft now uses across its Copilot products. Microsoft says it works with 20-plus frontier models plus bring-your-own-key, and developers can run it through Microsoft Foundry or an SDK.

Routing is the product now

Yes, a router sounds like a polite way to hand you cheaper models. GitHub's numbers argue the other way, and they are GitHub's numbers.

GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 Max. It says Fusion got there at about half the average API cost per task, with about half the tokens and about 40% less time per task.

Half the cost.

If that holds on your code, hand-picking the biggest model for every prompt gets hard to defend. The agent harness around the model now decides much of what you pay and what you get. Microsoft and GitHub shipping a router and a harness on the same day says they think so too.

One caution. A coding index is a set of benchmark tasks, not your repository. Treat the score as a reason to test, not a result.

What it costs and who controls it

Billing is per model. GitHub says you pay each model's standard rate for whatever Fusion calls. For a limited time, paid Copilot plans get a 30% discount on model costs inside Fusion. The changelog gives no end date.

Fusion runs in VS Code, the GitHub Copilot app, the Copilot CLI and JetBrains IDEs, with more clients promised. To try it, update your client, open the model picker, choose Copilot Fusion and pick an effort level.

For companies, Fusion uses only the models enabled in enterprise and organization policies. Admins can see which models it chooses from, switch a model off for Fusion alone, and set Fusion at any effort level as the default for their users.

That last control is the big one. An admin can make routing the default for a whole engineering team in one setting.

GitHub does not publish the full list of models Fusion can call, or any usage limits.

Test it before the discount ends

Run Fusion Standard on a week of your normal tasks and compare the bill with your usual model. Then try Max on the one task that keeps failing. If you want a reasoning effort setting to map this onto, this is the same idea one level up: you set how hard to try, and the system decides how.

Watch for the end date on that 30%. Then watch your invoice.