GitHub Copilot Fusion is now generally available as of October 28, and the benchmark headline is hard to ignore: Fusion Max scored within two points of Claude Code on the Artificial Analysis Coding Agent Index at roughly half the cost per task. Before you reorganize your entire workflow around that number, there is something you need to know. The 30% introductory discount on Fusion model costs expires November 30. That benchmark was taken with the discount applied.
What Fusion Actually Does #
The idea behind Fusion is that you stop picking a model and start picking an effort level. You choose Standard, High, or Max — and Fusion figures out the rest. Under the hood it runs three execution patterns depending on what the task needs.
Single: One model handles everything. Used for straightforward tasks where added model calls aren’t justified.
Cascade: An efficient model produces the first draft. A judge model evaluates it and either accepts or escalates to a stronger model. Cost only increases when the first attempt falls short.
Critique: One model drafts while multiple models critique in parallel. The original model revises based on that feedback. Reserved for demanding reasoning tasks that benefit from multiple perspectives.
GitHub says Fusion only adds model calls when a task warrants them — which is the right claim to make, and also one that’s difficult to verify until the December bill arrives.
The Four Tiers #
Fusion Standard prioritizes balanced cost, quality, and latency. Routine tasks typically route to a single model. This is also the tier where local model routing applies.
Fusion High handles substantial multi-step tasks with moderate token savings while routing to more capable models when needed.
Fusion Max routes to the most capable available models and uses roughly 7.2 million tokens per task on average. It produced the benchmark numbers dominating coverage this week.
Fusion Fast powers inline chat in VS Code. Fast models only, optimized for low latency.
The Benchmark Numbers, Honestly #
| Tool | Score | Cost/Task (now) | Cost/Task (Dec 1+) |
|---|---|---|---|
| Fusion Max | 60 | $5.59 | ~$7.99 |
| Devin Fusion | 62 | $8.00 | $8.00 |
| Claude Code (Opus 5.5) | 61.7 | $11.42 | $11.42 |
The two-point gap between Fusion Max and Claude Code is small enough that most teams won’t notice it in practice. The cost advantage is real. What the coverage has mostly glossed over is that the $5.59 figure is a discounted price. Without the 30% promotional discount, Fusion Max runs closer to $7.99 per task — still cheaper and faster than Claude Code, but the margin is narrower than the launch headlines suggest.
For reference: Devin Fusion scored 62 at $8.00 per task, with no promotional discount. The competitive picture is tighter than it looks.
Local Model Routing: The Underreported Cost Lever #
Fusion Standard now supports local model routing in public preview. Two models are available: MAI Code 1.1 Flash from Microsoft, and Qwen3.8 Flash Next. On Windows with supported hardware, MAI Code 1.1 Flash downloads and activates automatically when you select Standard tier. Qwen3.8 Flash Next can be configured manually on any platform via a local OpenAI-compatible server.
Enabling it in VS Code requires one settings change:
{
"chat.copilot.localModel.enabled": true
}
After that, selecting Standard in the Fusion panel lets the system route between local and cloud models automatically. Hardware requirements are significant — 96 GB of combined RAM and 60 GB of storage for MAI Code 1.1 Flash — but for teams that can meet them, running routine tasks locally at zero token cost is a real offset against December billing.
The December Cliff #
The 30% discount ends November 30. This has happened before. On September 1, GitHub’s earlier promotional credit period expired and Enterprise credit pools shrank 44% overnight. Business plans shrank 37%. Seat prices did not change. Most Enterprise admins had not seen their real bill during the subsidy window because the promotional credits had been absorbing overages.
GitHub Copilot’s shift to usage-based billing in June already generated persistent community complaints — one developer on Copilot Pro+ burned through 8% of their monthly allotment in two hours. Agentic workflows, which are exactly what Fusion Max is optimized for, are the most credit-intensive use case. Teams running parallel multi-repo Copilot Projects sessions will feel the December transition most directly.
Three Things to Do Before November 30 #
- Set organizational credit caps. GitHub’s admin dashboard supports per-user and per-organization credit limits. Set them based on your actual usage data from the past 30 days, not your included credit allotment. Fusion Max workflows will exceed the latter quickly.
- Audit which tasks actually need Max. File edits, explanation requests, and test generation for straightforward functions do not require Fusion Max. Move those to Standard or High. Reserve Max for complex architectural reasoning, large refactors, and multi-step debugging where the cascade and critique patterns earn their cost.
- Test local model routing on Standard. If your team is on Windows with hardware that meets requirements, MAI Code 1.1 Flash running locally has no per-token cost. That is a meaningful offset for routine tasks currently consuming cloud credits.
The Bigger Picture #
GitHub processed 7.38 billion commits in September 2026 — more than five times the count from a year earlier. Over 40% of pull requests in that month started with an AI agent’s first draft. The GitHub Universe 2026 announcement was explicit: “completed work, not more code, is the output that matters.”
That positioning required building at a different scale — 4 million CPU cores, 160 petabytes of storage, and 50 terabits per second of network capacity added this year, with multi-region Azure deployment arriving in early 2027. The platform ambition is credible. The pricing transition, like every pricing transition in this space, rewards teams that pay attention to the details before the bill does.