GitHub Copilot Fusion picks the models for you, at a 30% discount for now

GitHub has made Copilot Fusion generally available. Developers pick Fusion and an effort level, and Copilot chooses which models handle each prompt. It works on all Copilot plans, according to GitHub’s changelog.

Three effort levels and three ways to combine models

Fusion replaces the model picker choice with an effort setting: Standard for everyday coding, High for harder tasks that need deeper reasoning, and Max for the most demanding work. For every prompt, Copilot then decides which models to use and how they work together.

GitHub describes three workflows. In single, one model solves the task directly. In cascade, a cheaper model drafts an answer and a quality check either accepts it or escalates to a stronger model. In critique, one model drafts, a second reviews, and the first revises.

GitHub says Fusion combines two things it already had: the multi-model orchestration from Project HydraFusion, which ran as a research preview, and the model selection behind Copilot’s auto mode. Preview users who had turned on auto, HydraFusion or HydraFusion Max will see those choices moved to the matching Fusion level.

Where it runs and what it costs

Fusion is supported in Visual Studio Code, the GitHub Copilot app, GitHub Copilot CLI and JetBrains IDEs. GitHub says more clients are coming. To use it, update the client, open the model picker, select Copilot Fusion and choose a level.

Billing follows the models used: each one is charged at its standard rate. For a limited time, people on paid plans get a 30% discount on model costs when they use Fusion. GitHub has not said when that offer ends.

Administrators keep control. Fusion draws only on models enabled in enterprise and organization policies. Enterprise admins can see which models Fusion chooses from and disable one for Fusion without turning it off for everyone. They can also set Fusion, at any effort level, as the default through Enterprise managed settings.

GitHub also published a separate, opt-in experimental preview that lets Fusion route between cloud models and local ones. It covers MAI Code 1.1 Flash, a managed download on supported Windows devices, and bring-your-own models such as Qwen3.8 Flash Next through a local OpenAI-compatible server. GitHub’s changelog cautions that routing decisions are still made in the cloud, so a session is not offline, and that local models may have different safety protections than cloud ones.

Performance claims come from GitHub

GitHub’s headline figure comes from the Artificial Analysis Coding Agent Index, where Fusion Max scored 60. That is within two points of Claude Code running Opus 5.5 Max, according to GitHub. The company adds that Fusion needed about half the average API cost per task, about half the tokens and about 40% less time. These are GitHub’s numbers as published in its changelog. The post does not give the task mix or setup behind them, and they have not been independently verified here.

“Copilot Fusion eliminates that friction, optimizing model selection without requiring developers to manage it themselves.”

That is the quote GitHub attributes to Hari Pokhrel, a principal software engineer at AT&T.

A bet that routing, not a single model, is the product

The discount is the strategic clue. GitHub’s documentation already offers a 10% discount on model costs for paid-plan users of auto model selection, so the 30% Fusion offer is a clear push to move people from picking a model to letting Copilot pick. When the router decides, GitHub controls cost and quality, and the choice of model supplier matters less to the user.

The timing fits the load GitHub says it is carrying. Techmeme, citing The New Stack, reports that GitHub now sees 7.4 billion commits per month, up from 2.9 billion in August. Cheaper cascades that keep strong models for hard prompts are one way to serve that kind of growth, though GitHub’s post does not say that is the aim.

It also sits beside Microsoft’s other moves this week. We reported that Microsoft is selling the harness behind GitHub Copilot as a managed Foundry service, and that the Copilot app added multi-repo Projects and a computer use mode. Together they point to Copilot being sold as an orchestration layer over several models, not as one model.

Open questions remain. GitHub does not list which models Fusion picks from at each level, and the preview-to-general-availability change means earlier HydraFusion results may not match what ships. Developers who want a fixed model can keep choosing one in the picker.