GitHub ships Copilot Fusion, a router that picks AI models per task

Overview

Updated 1 hour ago

For two years, using an AI coding assistant meant picking a model from a dropdown and living with that choice for the whole session. GitHub's new Copilot Fusion removes the dropdown. You pick an effort level, and the system decides which models to call, in what order, and whether a second model should check the first one's work.

GitHub says the approach reaches near-frontier coding scores at roughly half the cost and 40% less time per task. Independent testing supports the cost and token savings but shows Fusion trailing the best single model on repository-editing tasks. The question now is whether enterprises trust a router enough to make it the default.

Why it matters

If routing holds up, the model you pick matters less than the system that picks it, and coding agents get cheaper per task.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

Fusion Max score on the Coding Agent Index

Artificial Analysis scored Fusion Max at 60, about two points behind Claude Code with Opus 5.5 Max.

$5.59

Average cost per task

Artificial Analysis measured $5.59 per task for Fusion Max, against $11.42 for Claude Code with Opus 5.5.

7.2M

Tokens per task

Fusion used about half the tokens of the Opus 5.5 Claude Code configuration in the same test.

40%

GitHub-reported time reduction

GitHub says Fusion finishes tasks about 40% faster; the underlying per-task timing data has not been published independently.

30%

Limited-time Fusion discount

Paid Copilot plans receive 30% off model costs when using Fusion for a limited period.

Voices

Curated perspectives — historical figures and your fellow readers.

People Involved

Jay Parikh

Executive vice president of CoreAI at Microsoft

Leading Microsoft's AI platform strategy

Organizations Involved

GitHub

Software platform, subsidiary of Microsoft

Operator of Copilot and publisher of Copilot Fusion

GitHub builds the Copilot coding assistant and sets which models it can route to.

Microsoft

Technology company

Parent company of GitHub

Microsoft owns GitHub and sells Copilot through its enterprise agreements.

Artificial Analysis

Independent AI benchmarking firm

Published the Coding Agent Index comparison

Artificial Analysis runs independent benchmarks of AI models and agents.

Anthropic

AI model developer

Maker of Claude and the Opus 5.5 model used as the comparison point

Anthropic makes the Claude models that Fusion can route to and that benchmarks compare against.

Timeline

September 2026 November 2026

4 events Latest: Yesterday

Tap a bar to jump to that date

November 2026
Fusion available across Copilot plans and clients
Latest Rollout

The system is available in VS Code, the Copilot app, Copilot CLI and JetBrains IDEs, with administrator controls over which models it can use.
Yesterday
October 2026
GitHub announces Copilot Fusion general availability
Product launch

Fusion combines HydraFusion's orchestration with Copilot's automatic model selection and adds three effort levels.
5 days ago
Independent benchmark shows cost gap, quality trade-off
Benchmark

Artificial Analysis scores Fusion Max at 60 and $5.59 per task, against about 62 and $11.42 for Claude Code with Opus 5.5 Max.
5 days ago
September 2026
HydraFusion preview arrives in VS Code
Product

GitHub's earlier multi-model orchestration system appears in the model picker for VS Code and the Copilot app.
September 30th, 2026

Scenarios

Fusion becomes the default for enterprise Copilot users

Possible Resolves by Apr 30, 2027

Discussed by: GitHub's changelog and enterprise administrator documentation

Administrators can set Fusion as the default at any effort level through enterprise managed settings. If large organizations take that option, most Copilot users stop choosing models at all, and GitHub's router becomes the main gatekeeper for which model providers get enterprise traffic.

Independent tests confirm the cost gap holds

Likely Resolves by Q2 2027

Discussed by: Artificial Analysis

The first independent benchmark showed Fusion at roughly half the per-task cost of Claude Code with Opus 5.5 Max. A follow-up index would show whether that gap survives new models and different task mixes, or whether it was specific to the October test set.

A rival ships its own model router

Possible Resolves by Q2 2027

Discussed by: Technology trade coverage of the Fusion launch

Fusion's premise is that routing beats picking one model. If that premise holds, Anthropic, OpenAI or another coding-tool vendor has reason to ship a comparable router rather than compete on single-model benchmarks. The trigger would be a rival announcing multi-model orchestration inside its own coding product.

Repository-task gap limits Fusion adoption

Uncertain Resolves by Q2 2027

Discussed by: Artificial Analysis benchmark results

Fusion scored 51 on FrontierCode, a repository-change evaluation, against 56 for Claude Code with Opus 5.5 and 54 for Devin Fusion. If that gap persists, teams doing large codebase edits may keep choosing a single strong model, and Fusion's savings may stay concentrated in terminal and question-answering work.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

1990s-2000s

Ensemble methods in machine learning (1990s-2000s)

Statisticians and computer scientists showed that combining many weaker models often beat a single strong one. AdaBoost arrived in 1995, and Leo Breiman's Random Forests paper followed in 2001. The idea was counterintuitive at the time: more models, not a better model, produced the gain.

Then

Ensembles became standard in competitions and production systems.

Now

The approach established that orchestration can substitute for raw model strength, which is the premise Fusion applies at runtime.

Why this matters now

Fusion is an ensemble method for coding tasks, choosing at request time which models to combine rather than training them together.

1957-1960s

Optimizing compilers replace hand-written assembly (1957-1960s)

IBM's FORTRAN compiler in 1957 generated machine code that programmers had previously written by hand. Early compilers produced worse code than expert assembly writers, but they improved quickly and removed a decision from the programmer's plate.

Then

Programmers kept hand-tuning performance-critical routines while using compilers elsewhere.

Now

Compilers became the default, and hand-written assembly shrank to a niche.

Why this matters now

Fusion asks developers to stop choosing a model the way early programmers stopped choosing instructions. The trade-off is the same: convenience and cost against peak quality on hard cases.

2006-2010s

Cloud auto-scaling abstracts server sizing (2006-2010s)

Amazon Web Services launched Elastic Compute Cloud in 2006, and auto-scaling tools followed. Instead of picking a server size in advance, teams set a policy and let the platform add or remove capacity.

Then

Teams saved money on idle capacity but lost fine control over performance.

Now

Managed scaling became the norm, and cloud providers gained leverage over infrastructure choices.

Why this matters now

Fusion moves model selection from the developer to the platform, the same shift auto-scaling made for servers. It also gives GitHub similar gatekeeping power over which models get used.