What GitHub Copilot Fusion means for your model choice

GitHub has made Copilot Fusion generally available, according to the GitHub Blog changelog. Instead of choosing one model from Copilot’s model picker, you choose Fusion and an effort level. Copilot then decides which models handle each prompt. One detail decides what this means for you: billing. According to GitHub, you are billed for each model Fusion uses, at that model’s standard rate. Paid plans get a 30% discount on model costs for a limited time.

The background you need

The release note assumes you know a few terms. Here they are.

The model picker. Copilot’s chat and agent features in editors like Visual Studio Code have a dropdown. It sets which large language model answers you. Models trade off capability, speed, and price. Larger models usually handle harder reasoning better, but they cost more and respond more slowly. Smaller models are fast and cheap, and they are often good enough for routine edits.

Tokens. Models read and write text in small chunks called tokens. Usage-based AI pricing is usually charged per token, and each model has its own rate. So the same task can cost very different amounts. The total depends on which model runs it and how many tokens that model consumes.

Routing versus orchestration. Routing means picking one model for a given job. Orchestration means coordinating several models on one task. For example, one model might plan the work while another writes the code. GitHub says Fusion does both.

Where Fusion came from. Two features came before it. Copilot auto model selection let Copilot pick a model for you. GitHub describes its selection as fine-tuned. Project HydraFusion was GitHub’s multi-model orchestration effort. According to GitHub, Fusion merges the two. It uses auto selection’s model choice and HydraFusion’s orchestration to balance performance and cost. The release note tells research preview participants that their earlier auto, HydraFusion, or HydraFusion Max selections will carry over. So these were options people could already turn on before general availability.

Effort level. This setting tells the system how hard to work on a prompt. The source names two Fusion tiers: Max, in GitHub’s benchmark, and Standard, in a customer quote. It does not list every level or describe how they differ.

A useful analogy. Think of Fusion as a general contractor. You don’t hire the electrician and the plumber yourself. You hire the contractor, who brings in whichever trades the job needs. You then pay for the trades that showed up. The analogy breaks in one place. A contractor usually gives you a quote first. The source does not describe any cost estimate before a prompt runs. You set an effort level, not a budget.

What actually changed

Some of this is new. Some of it was already true.

Already true: automatic model selection existed as Copilot auto. Multi-model orchestration existed in Project HydraFusion. Picking a specific model yourself is still part of the picker. GitHub presents Fusion as an alternative to picking a model for each task.

New, according to GitHub:

• General availability. Fusion is available on all Copilot plans. It works in VS Code, the GitHub Copilot app, GitHub Copilot CLI, and JetBrains IDEs. GitHub says support for more clients is coming soon. You need to update your client to see it.
• One entry instead of several. Auto selection and HydraFusion are now one picker option with effort levels. In GitHub’s words, when you select Fusion “you aren’t selecting a specific model.” For each prompt, Fusion decides which models to use and how they work together.
• Preview migration. If you used auto, HydraFusion, or HydraFusion Max in the preview, your setting moves to the matching Fusion effort level automatically.
• Admin controls. Administrators control which models Fusion can use and whether Fusion is the default.
• Billing terms. Each model Fusion uses is billed at its standard rate. Paid plans get a limited-time 30% discount on model costs when using Fusion.
• Performance claims. GitHub says it has improved performance since Project HydraFusion launched. On the Artificial Analysis Coding Agent Index, which scores coding agents, GitHub reports that Fusion Max scored 60. That is within two points of Claude Code with Opus 5.5 Max. GitHub says Fusion Max reached that score at about half the average API cost per task. It also used about half as many tokens and took about 40% less time per task.

What it means for developers who don’t follow every release note

Start with billing, because it shapes everything else.

Fusion has no flat price. A prompt costs whatever the models Fusion picked cost, at their normal rates. That means two things are outside your direct control. One is which models run. The other is how many of them run on a given prompt. Your main control is the effort level. You also inherit whatever limits your administrator set on which models Fusion may use.

Compare that with picking a model yourself. When you choose one model, you know which rate applies, even if you can’t predict the token count. With Fusion, the rate itself changes from prompt to prompt. This can work in your favor if Fusion sends simple tasks to cheaper models. The AT&T quote in the release note describes exactly this pattern. A Principal Software Engineer there says developers often chose a more powerful model for a complex task and then forgot to switch back for everyday work. He says Fusion Standard improved token efficiency across AT&T. That is one customer’s account, published by GitHub.

Read the benchmark carefully. GitHub compared Fusion Max with Claude Code using Opus 5.5 Max. That is a different product in a specific configuration. It does not compare Fusion with you choosing models inside Copilot. So it tells you how Fusion Max stacks up against that one setup on that index. It does not tell you whether Fusion will cost less than your own model choices.

The discount also matters for timing. GitHub says the 30% discount is for a limited time. If you judge Fusion’s cost during the discount, the numbers may look different once it ends.

What to watch next:

• Your Copilot usage and billing in the weeks after you switch, compared with before.
• Any announcement of when the discount ends.
• Whether your organization makes Fusion the default, and which models it allows.
• Support for more clients, if your editor is not on the current list.

What we don’t know yet

The release note leaves several questions open:

• How long the discount lasts. GitHub says only that it is for a limited time.
• The full set of effort levels. The source names Max and Standard. It does not list all levels or explain how each one changes model choice or cost.
• How routing decisions are made. GitHub says Fusion analyzes each task. It does not explain the criteria.
• Whether you can see which models ran. The source does not say whether the product shows you which models a prompt used, or what each one cost.
• How Fusion compares with manual selection. The published benchmark uses Claude Code with Opus 5.5 Max as its comparison point. No comparison with hand-picked Copilot models is given.
• Which clients come next. GitHub says additional clients are coming soon but does not name them.