AI & DEV NEWS · · 4 min read

GitHub Copilot Fusion is GA: one picker, several models, a bill per model

GitHub made Copilot Fusion generally available on October 28. You pick an effort level, and Copilot routes each prompt to one or more models. github.blog claims near-frontier results at about half the cost.

How GitHub Copilot Fusion picks models for each prompt

In the model picker you choose Copilot Fusion, then an effort level: Standard for everyday coding, High for harder tasks, Max for the most demanding ones. After that, Copilot decides which models to use and how they work together. According to the GitHub changelog, it has three workflows.

Single means one model solves the task. Cascade means a cheaper model drafts, a quality check decides whether to accept, and a stronger model takes over if not. Critique means one model drafts, another reviews, and the first revises. It combines the earlier Project HydraFusion research preview with Copilot's auto model selection, and old Auto and HydraFusion selections carry over to the matching Fusion levels.

It runs in VS Code, the GitHub Copilot app, Copilot CLI and JetBrains IDEs, on all Copilot plans. More clients are promised, with no dates given.

The benchmark number is GitHub's, and it measures one thing

Here is the headline. On the Artificial Analysis Coding Agent Index, GitHub says Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max. It also says it got there at about half the average API cost per task, about half the tokens and about 40% less time. All of that comes from GitHub's own changelog. We haven't seen an independent rerun yet.

One index is one index. It tells you nothing about your monorepo, your test suite or your review habits. The changelog also gives no absolute dollar figure, so 'half' is relative to a baseline you can't see.

One more caution, and this is our reasoning, not GitHub's data. Multi-model setups don't save money by default. Every extra draft, review and revision is another billed call, so Cascade only pays off if the cheap model's drafts get accepted often enough. The changelog doesn't say how often that happens.

One picker, but the bill still has several lines

Fusion bills each model it uses at that model's standard rate. For a limited time, paid plans get a 30% discount on model costs when using it. The changelog gives no end date, so assume the discount will disappear and judge Fusion at full price.

Critique and Cascade mean one prompt can touch more than one model. Spend per task will vary with how the router behaves, not just with how big your prompt is. That's a different shape from picking one model and watching a flat per-request cost. aimadetools.com makes the same point: billing stays tied to individual models, not a flat price per agent task.

AT&T's Hari Pokhrel says in the changelog that Fusion Standard cut unnecessary spend. People used to pick a powerful model for a hard task and forget to switch back. That's a plausible failure mode, and we've seen versions of it. It's still a customer quote on the vendor's page.

Admin controls are the part that matters at scale

Fusion only uses models enabled in your enterprise and organization policies. Enterprise admins can see which models it picks from. They can disable a model for Fusion without disabling it for everyone, and they can set Fusion at any effort level as the default through managed settings.

That helps with the boring problem: nobody has to remember which dropdown entry is the expensive one. It also means a model policy change now affects routing, not only the picker.

What we would do this week

Nothing dramatic. If your team is on Copilot, run Fusion Standard as the default for a small group and keep a manual model choice for everyone else. Compare accepted changes, retries and review time over a week or two, at full price and not the discounted one.

Admins should check the model list Fusion can draw from before turning it on as a default. If you build automations outside the IDE, Fusion doesn't help. Routing there is still your job, in an n8n flow or a model gateway. See our AI workflow automation page for how we think about that layer.

  • Pilot with one team, log accepted results, retries and review effort.
  • Review enabled models in enterprise policy before setting a default.
  • Re-check cost when the 30% discount ends.

Sources

  1. Route and orchestrate models with GitHub Copilot Fusion · GitHub Changelog
  2. AI Dev Weekly #32: GPT-6.1 Terra, Gemini Carbon, Kimi K3.1 and Copilot Fusion · AI Made Tools

Want this working in your stack, not just in a demo?

We are two senior engineers who build ai workflow automation with a fixed price per written scope.

Keep reading.