This GitHub Copilot Fusion review looks at what GitHub shipped on October 28, 2026, what is confirmed so far, and what is not. GitHub made Copilot Fusion, its multi-model routing and orchestration system, generally available on October 28, 2026, at GitHub Universe 2026. Developers can now choose Fusion instead of a single model and let Copilot decide which models do the work.
How this article was made: it is based on official announcements, documentation and the sources listed at the end, not on our own hands-on testing. See our editorial policy.
Table of Contents
What Copilot Fusion is
GitHub’s documentation describes Copilot Fusion as an intelligent runtime model router and orchestrator. You select it in the model picker, but you aren’t selecting a specific model.
For each prompt, Fusion determines which models to use and how they work together. It combines the multi-model orchestration of Project HydraFusion with the fine-tuned model selection behind Copilot’s auto model selection. Project HydraFusion was introduced as a research preview in September.
Who it is for in this GitHub Copilot Fusion review
Fusion is aimed at developers who are tired of choosing a model for every task. It suits teams that want one default setting, and enterprise managed settings can set Fusion, at any effort level, as the default model for Copilot users.
If you already pick models by hand for specific jobs, such as a heavy model for architecture work and a cheap one for boilerplate, Fusion tries to automate that decision. The sources I found for this GitHub Copilot Fusion review do not publish enough detail to say how well it makes those choices.
Key features
Routing based on task needs
GitHub’s Parikh wrote that Copilot Fusion evaluates what a task needs across reasoning, code generation, debugging, and tool use, then chooses how to execute across one or more models. The docs say it only adds model calls when a task warrants them. That is meant to avoid unnecessary costs and delays on simpler tasks.
Three effort levels
Fusion has three effort tiers: Standard, High, and Max. According to GitHub, the flagship Max tier improves task-completion rates over top-tier frontier baselines while capping estimated inference cost at 1.1 times the baseline model. That is a GitHub claim, not an independent benchmark. The docs also include a table of which effort levels are available on which Copilot plan, so check it for your plan.
Visible progress
While Fusion works, you can follow the progress of each step and the model being used. That helps if you want to see where your spend goes.
Experimental local model routing
GitHub also announced experimental support for routing tasks between cloud and locally running AI models. Fusion can use an installed local model only from the Standard effort level. In the Copilot app, you turn on “Local model auto routing” in the Experimental section of the settings.
Where you can use it
Fusion is generally available across all GitHub Copilot plans in Visual Studio Code, GitHub Copilot CLI, the GitHub Copilot app, and JetBrains IDEs. More clients are planned. If none of the models Fusion uses are available to you, it doesn’t appear in the model picker.
Pricing confirmed so far
The billing rules in the docs are simple:
- You are billed for each model Fusion uses, at that model’s standard rate.
- There is no separate charge for Fusion itself.
- Through November 30, 2026, users on paid Copilot plans get a 30% discount on model costs when using it.
The discount is temporary. Any savings you see before December 1 will partly come from the promotion, not from the routing. For per-model rates, GitHub points to its “Models and pricing for GitHub Copilot” page.
Limitations and open questions
The headline claim of this article’s working title, that Fusion matches Claude Code at half the cost, is not confirmed by the sources I found. GitHub’s cost figure is a cap relative to a baseline model: estimated inference cost capped at 1.1 times the baseline model. That is not a “half the cost” claim, and I found no comparison against Claude Code.
Other points are also unknown or unverified:
- Independent benchmarks of task-completion rates are not available in the sources I read.
- The exact list of models Fusion routes to, and which it excludes, is in the docs. I did not find enough detail to reproduce it here.
- The docs have a limitations section. Read it before standardizing your team on Fusion.
- The local-model routing is experimental and limited to the Standard level.
Alternatives
If you want to choose models yourself, Copilot’s model picker still lets you. You can also use a dedicated terminal agent. Our guide to Claude Code for business covers that route, and Claude Code Mods shows how to customize it with TypeScript plugins. For help choosing between frontier models for agent work, see GPT-6 Sol vs Luna vs Claude Opus 5.5.
Verdict
This GitHub Copilot Fusion review comes down to one point: Fusion is a credible way to stop hand-picking models, and it adds no orchestration fee. It is also new, and its quality and cost claims come from GitHub. Treat the “matches Claude Code at half the cost” idea as unproven.
A reasonable approach is to run Fusion on a few real tasks while the 30% discount lasts. Compare the results and your billing against your current model choices before changing team defaults. Any final GitHub Copilot Fusion review should rest on your own repositories, not on launch-week numbers.