GitHub Copilot Fusion: what it is and how to use it

GitHub Copilot Fusion is a router built into Copilot. You stop choosing a model and choose an effort level instead (Standard, High or Max), and Copilot decides which Anthropic or OpenAI models handle each prompt. To use it, update VS Code, JetBrains, Copilot CLI or the Copilot app, open the model picker, select Copilot Fusion and pick a level.

The posts behind it

GitHub made Fusion generally available at Universe on Wednesday. The first number came from outside GitHub. @ArtificialAnlys posted its Coding Agent Index results, saying Fusion "orchestrates between Anthropic and OpenAI models" and lands 2 points behind Claude Code running Opus 5.5 at max effort. Its listed cost per task was $5.59 for Fusion and $11.42 for Claude Code. It also put Fusion first on Terminal-Bench 4.0 with a score of 71, using about half the tokens Opus 5.5 used in Claude Code.

About ninety minutes after that, the official @github account said Fusion "can coordinate multiple models and workflows when one model alone isn't enough," and announced a 30% discount on Fusion usage. GitHub's Kyle Daigle, posting as @kdaigle, claimed up to 77% lower cost and up to 40% faster runs, and added his own experience: "It's replaced choosing models for me over the last two months."

The benchmark comes from Artificial Analysis. The "up to" figures are GitHub's own, so read them as marketing.

How it works

The Fusion docs say Fusion picks a model pool and an execution pattern for every prompt, and the pattern can change within one session:

You only see the final answer. To find out which models did the work, hover over the effort level name under a response. GitHub doesn't publish a fixed model list, and the pool will change over time.

Each effort level suits a different kind of work. Fusion Standard is for everyday tasks and usually sends each one to a single model. Every plan gets it, Free, Student and Pro included. Fusion High is meant for complex bugs and multi-file changes, and Fusion Max reaches for the strongest models on the hardest reasoning problems. Both are limited to Pro+, Max, Business and Enterprise.

How to try it

1. Update your editor, the Copilot app or Copilot CLI.
2. Open the model picker and choose Copilot Fusion.
3. Start on Standard. Move to High for a gnarly bug that touches many files, and keep Max for the job you'd otherwise hand to Opus 5.5 at max effort.
4. Try a prompt like "Find why the checkout total is off by a cent on discounted orders, fix it, and add a test," then hover over the effort label to see what Fusion picked.

GitHub also shipped an experimental option that sends some Standard work to models on your own machine. In Copilot CLI, start with --experimental or run /experimental on. In the Copilot app, turn on Local model auto routing. The supported local models are MAI Code 1.1 Flash on qualifying Windows PCs and Qwen3.8 Flash Next through a local server such as Ollama. Read our post on whether Qwen3.8-Flash-Next is safe before you install that one. Routing decisions still happen in the cloud, so local models don't make Copilot work offline.

What it costs

Fusion has no fee of its own. You pay each model it calls at that model's standard rate, which means a Critique run on Max can burn more AI credits than one Standard call. Through November 30, 2026, paid plans get 30% off model costs when they use Fusion. For Business and Enterprise, admins control which models Fusion can reach, can block a model for Fusion alone, and can make a Fusion effort level the default for everyone.

Two limits from the docs are worth knowing. Fusion doesn't start subagents. And if Fusion throws out a draft, any file edits that draft already made stay in your workspace, so review the diff before you commit.

My read

I care less about the benchmark than about what Fusion does to the daily routine. Picking a model has turned into a chore of its own, with new releases every week, and Fusion hands that choice to GitHub. You can see a similar pull in Cursor's Agent SDK and in OpenAI's push on GPT-6.1 Terra: the model is becoming a setting you rarely touch, and vendors are competing on what a finished task costs.

For a business, that's the number to track. Measure AI spend per job done rather than per seat or per token, because once a router can mix several models on one request, per-token prices stop telling you much. If you want help working out which of your own processes are worth handing to an AI tool, and what each one would cost, New Face Design's free process audit is a good place to start.