Copilot Fusion: Model Routing for AI Coding Explained

GitHub Copilot Fusion is a setting that lets Copilot choose the AI models for you. GitHub announced on 28 October 2026 that it is generally available, meaning it is open to everyone and no longer a research preview. Instead of picking one model, you pick an effort level, and Copilot decides which models to use.

This matters if you pay for AI coding by usage, or if nobody on your team wants to keep up with a model list that changes every month. Below: what changed, how it works, what it costs, and what to check before you rely on it.

What is GitHub Copilot Fusion?

Until now, most AI coding tools asked you to choose a model: a bigger, slower, pricier one for hard work, or a smaller, cheaper one for easy work. Many people guessed, or left the same model on for everything.

Fusion takes that choice away. In the model picker you select Copilot Fusion and one of three effort levels. GitHub's announcement describes them like this:

• Standard: favours versatile models for everyday coding.
• High: favours capable models for challenging tasks and deeper reasoning.
• Max: favours the most capable models for the most demanding tasks and complex reasoning.

GitHub says Fusion combines the multi-model orchestration of its earlier research project, HydraFusion, with the model selection already in Copilot's "auto" option. People who used auto, HydraFusion or HydraFusion Max in the research preview have been moved to the matching Fusion effort level automatically.

How does Fusion decide what to do?

Fusion uses one of three patterns. Think of them as three ways of staffing a job:

• Single: one model handles the task directly.
• Cascade: an efficient model writes a first answer. A quality check then either accepts it or passes it to a stronger model. It is like a junior colleague drafting and a senior one stepping in only when needed.
• Critique: one model drafts, a second model reviews the draft, and the first model revises. It is like a writer and an editor.

You do not choose the pattern; Fusion does. The idea is that easy work should not pay for the strongest model, and hard work should not be left to a weak one.

What does it cost?

This is the part to read carefully. GitHub says you pay the standard rate of each model Fusion uses. So if a task is drafted by a cheaper model and then reviewed by a pricier one, you pay for both.

Two consequences:

• Cost per task is less predictable. It depends on the choices Fusion makes, not on one model you picked.
• There is a temporary discount. For a limited time, people on paid Copilot plans get 30% off model costs when using Fusion. GitHub's announcement says only "limited time", so do not build a budget on it.

Is it really cheaper and just as good?

GitHub reports that Fusion Max scored 60 on the Artificial Analysis Coding Agent Index. That is within two points of Claude Code running Opus 5.5 at its Max setting. GitHub also says Fusion Max did this at about half the average API cost per task (API cost is what you pay for each request to a model), using about half the tokens (the small chunks of text a model reads and writes) and about 40% less time per task.

Treat this as a claim to test, not a fact about your work. These are GitHub's own reported figures. A benchmark is a set of standard tasks; your codebase, your rules and your tolerance for mistakes are different. A fair test is simple: give the same ten real tasks to your current setup and to Fusion, then compare the results, the cost and the time.

Who can use it, and who controls it?

• Plans: all Copilot plans.
• Where: Visual Studio Code, the GitHub Copilot app, Copilot CLI (the command-line version, used by typing in a terminal window) and JetBrains code editors. GitHub says support for more clients is coming.
• How: update your client, open the model picker, choose Copilot Fusion, then an effort level.

For managers there are controls. Fusion only uses models that your enterprise and organisation policies already allow. Enterprise admins can see which models Fusion chooses from, and can switch off a model for Fusion alone without turning it off elsewhere. Admins can also make Fusion, at any effort level, the default through managed settings.

If you care about which AI providers handle your code, check the allowed-model list before you switch Fusion on for a team.

What should a team do this week?

1. Decide who owns the cost. Because the bill follows the models chosen, someone should watch the first few weeks of spending, not just the first day.
2. Run a small side-by-side test. Ten real tasks, same instructions, compare. Include one or two hard ones.
3. Check your model policy. Fusion respects it, but make sure the policy says what you intend.
4. Keep your own checks. A smarter model choice does not replace tests and review, whichever model wrote the change.
5. Ignore the discount when comparing. Compare at full price so the result still holds when the discount ends.

How does Atlas handle model choice?

Atlas is a team of AI engineers built on Claude Code, and it works differently from Copilot Fusion. Each engineer is its own Claude Code session with a written role, and the model and effort are tuned to that role. A design role and a testing role can therefore run on different settings, chosen once for the role.

What Atlas does not do today is change the model inside a single task, for example drafting with a cheaper model and escalating when a check fails. That is coming to Atlas, built in, so it will need no setup on your side. Until then, the per-role settings are what Atlas gives you. Atlas runs on your own Claude plan, so usage counts against that plan, and an optional daily budget warns at 50% and 80% and pauses the team at 100%. You can learn more at the Atlas home page.

Conclusion

Copilot Fusion is a real step towards "tell the tool how hard the job is and let it choose the model". It could save money and spare teams the model guessing game. But the bill follows its choices, the best numbers are GitHub's own, and the discount is temporary. Test it on your own work, give someone the cost, and keep your normal checks in place.