Copilot Fusion Now Generally Available — Does Auto Model Selection Really Cut Costs in Half?
機械翻訳 / Machine-translated
A
AI Friends
8時間前・15分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Have you ever thought, "I have no idea which AI model to pick," or worried about runaway bills from always choosing the most expensive model? Copilot Fusion, officially released by GitHub on October 28, 2026, is designed to solve exactly these problems. This article walks you through how it works, what it costs, and what to watch out for.
At GitHub Universe 2026 in San Francisco, GitHub announced the general availability of Copilot Fusion.
Previously, developers had to choose their AI model manually. With Fusion, you simply select "Copilot Fusion" and Copilot decides which model to use and how to combine models for each task.
The feature is built on two earlier initiatives: Project HydraFusion (an experiment in coordinating multiple models), which had been available in preview, and Copilot's automatic model selection. Fusion brings both together into a single experience.
Fusion first estimates the capabilities a task requires — reasoning, code generation, debugging, tool use, and so on — and then decides on an execution pattern.
Single means one selected model handles the task entirely on its own.
Cascade has an efficient model produce a solution, accepts it if it passes a quality check, and escalates to a higher-performing model if it doesn't.
Critique has one model produce a draft, which is then independently reviewed by a different model family. The drafting model is given one chance to revise based on the feedback.
GitHub's APAC VP compared Critique to "not auditing your own tax return yourself" (as reported by ITWire).
What users choose is a load level, not a model name.
There is also Fusion Fast, which is selected automatically when speed is the priority — users cannot specify it directly (ITmedia NEWS).
Here are the figures GitHub has published:
On Artificial Analysis's Coding Agent Index, Fusion Max scored 60 points — within 2 points of Claude Code (Opus 5.5 Max), according to GitHub.
Additionally, average API cost per task and total token usage are said to be roughly half, and time to completion is said to be about 40% shorter.
Keep in mind that these are GitHub's own figures, not independently verified results. It's worth testing on your own projects to see what holds up in practice.
In this announcement, GitHub only directly compares Fusion against Claude Code. No comparisons with other tools have been published, as far as we can confirm.
Here, then, is a comparison between Fusion and the traditional approach of selecting models manually:
For example, using a top-tier model to fix a small typo is wasteful. With Cascade, a lightweight model tries first, and if it passes the quality check, the task is done.
Fusion itself carries no fixed additional charge. You are billed the standard rate for each model Fusion uses.
Paid plan users receive a 30% discount on model costs when using Fusion. According to ITWire, this runs through the end of November. SmartScope reports the end date as November 30, 2026.
Information about which plans are supported varies by source. GitHub's official announcement states it is "available on all Copilot plans." However, SmartScope reports that High and Max require Pro+ or above. Please check GitHub Docs before subscribing.
Supported clients include Visual Studio Code, the GitHub Copilot app, Copilot CLI, and JetBrains IDEs. Additional clients are planned for the future.
No Japan-specific pricing or availability conditions have been announced, as far as we can confirm. That said, Visual Studio Code and JetBrains are both among the supported IDEs — both widely used in Japanese development environments, so the barrier to adoption should be relatively low.
Consider, for instance, a development lead at a software contracting firm. If team members each choose their own models, monthly AI costs can vary unpredictably. Setting Fusion as the default makes it easier to smooth out inconsistencies in usage.
For enterprise administrators, Fusion can only use the models enabled under your model policy. You can even exclude specific models from Fusion while keeping them available elsewhere — a structure that works well for teams with security review requirements.
Enterprise managed settings also let administrators set any load level as the default model for users.
Getting started is straightforward: update your client to the latest version, select Copilot Fusion in the model picker, and choose a load level.
Users who were using auto, HydraFusion, or HydraFusion Max during the research preview will be automatically migrated to the corresponding level.
A few caveats: because High and Max combine multiple models, they may consume more credits than Standard. It's also unclear how the discount will be handled after November 30.
For individual developers, a good approach is to start with Standard for everyday work and step up to High or Max only for heavier tasks — that way you can keep a close eye on costs.
A. The official announcement says all plans are eligible. However, standard per-model charges still apply. Paid plan users get a 30% discount through the end of November.
A. You choose the load level. GitHub's announcement does not specify which models are used internally.
A. GitHub says Fusion Max scores within 2 points of Claude Code (Opus 5.5 Max). These are GitHub's own numbers, so verify with your actual workloads.
A. Yes. Only models enabled in your model policy are eligible. You can also exclude specific models from Fusion specifically.
Start by selecting Standard in your usual editor and tracking your costs and experience over the course of a week.
This article is a cross-post from AI Friends.
