Copilot Fusion now picks the model for each task

How Copilot Fusion picks and combines models

Copilot Fusion shows up in the model picker next to individual models, but it is not a model. It is a routing layer that runs at request time. Copilot Fusion weighs each task against signals for reasoning, code generation, debugging and tool use, checks system health, and then returns a single response from the models it chose.

Fusion merges two things GitHub had shipped separately: Copilot's auto model selection and Project HydraFusion, a research preview that chained several models together. If you had auto, HydraFusion or HydraFusion Max turned on during the preview, your setting has moved to the matching Fusion effort level.

GitHub says more clients are coming soon. For now, Fusion works in these four:

• Visual Studio Code
• the GitHub Copilot app
• GitHub Copilot CLI
• JetBrains IDEs

Fusion switches models only at cache boundaries

Fusion can change models partway through a task, but only at natural cache boundaries or when the conversation gets much more or less complex. The cache is what lets a model reuse context it has already read instead of paying to process it again.

Keeping the main conversation on one model lets it reuse cached tokens, while reviewer models receive only the context they need, so switching rarely adds cache costs. GitHub built it this way so you don't pay for a model switch that wouldn't improve the answer.

You can see which models answered each prompt

While Fusion works, each step shows the model in use, and hovering over the effort level name under the response lists every model involved. Intermediate drafts can be revised or thrown away, so you only see the final answer.

Fusion has no context window of its own. The number shown in the picker is a conservative value based on the smallest limit among the models Fusion can call.

Three effort levels, with High and Max on Pro+ and up

After you select Fusion, you choose an effort level instead of a model. The level tells Fusion how to trade off quality, latency and cost.

Standard is meant for everyday work and uses one model per task. High targets substantial but well-scoped jobs, such as fixing a complex bug or making a change across several files. Max is for work that needs the deepest reasoning and the most thorough analysis.

Not every plan gets all three. Free, Student and Pro users get Standard only, while High, built for complex bugs and multi-file changes, and Max, built for the deepest reasoning, require Pro+, Max, Business or Enterprise. A fourth level, Fast, cannot be selected at all; it only runs inside Visual Studio Code's inline chat.

Fusion effort levels by Copilot plan (GitHub Docs)

Plan	Standard	High	Max
Free	Yes	No	No
Student	Yes	No	No
Pro	Yes	No	No
Pro+	Yes	Yes	Yes
Max	Yes	Yes	Yes
Business	Yes	Yes	Yes
Enterprise	Yes	Yes	Yes

Fusion picks Single, Cascade or Critique for each prompt

Based on the effort level and the task, Fusion runs each prompt through one of three execution patterns. Users never choose the pattern themselves, so one prompt in a session can be solved by a single model and the next can go through parallel critiques. Simple tasks skip the review step entirely, so they don't wait or spend AI credits on a review they don't need.

Cascade and Critique review a draft before answering

The three patterns differ in whether a draft gets checked and by how many models. Single sends the task to one model, Cascade lets a judge accept an efficient model's draft or escalate it, and Critique has several models critique the draft in parallel before the original model revises it.

Fusion's three execution patterns (GitHub Docs)

Pattern	Who drafts	Who reviews
Single	One model solves the task	No review
Cascade	An efficient model	A judge accepts the draft or escalates it
Critique	One model	Several models critique in parallel; the drafter revises

Reviews add time, and discarded drafts keep their edits

Picking a pattern adds very little latency on its own. Cascade and Critique take longer because of their review calls, and any file edits a discarded draft already made in your workspace stay in place until you undo them.

That matters most for Cascade: if a draft gets escalated to a stronger model, whatever that draft wrote to disk is still there. GitHub's advice is to review changes before you commit them. Fusion also runs separately from Copilot's subagents and never starts one.

Fusion Max scored within two points of Claude Code

GitHub backed the launch with a third-party benchmark, the Coding Agent Index from Artificial Analysis, which scores coding agents end to end. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index while using about half the average API cost and tokens per task and about 40% less time than Claude Code with Opus 5.5 Max. The two setups landed within two points of each other.

These are GitHub's own figures from the announcement. GitHub also quoted Hari Pokhrel, a Principal Software Engineer at AT&T, who said Fusion Standard improved token efficiency across the company by picking the right model for each task automatically.

Billing at each model's rate, 30% off through November 30

Fusion bills you for every model it calls, at that model's standard rate. There is no separate charge for Fusion, but High and Max can call more than one model per task, so the same task can consume more AI credits than it would on Standard.

Paid plans also get a limited-time discount: through November 30, 2026, model costs are 30% lower when they run through Fusion. AI credits are the same unit GitHub charges for other agent features, including GitHub agentic autofix is now generally available, which spends them on every run.

Admins can exclude models from Fusion only

Fusion only draws on models that your enterprise and organization policies already allow. An enterprise admin can stop Fusion from using a model while leaving that model available to anyone who selects it manually. Admins can also see the full list of models Fusion picks from and, through Enterprise managed settings, make Fusion the default model at any effort level.

A few other rules shrink the pool. Fusion skips any model that is:

• not included in your plan
• blocked by policies that limit Copilot to data-resident or FedRAMP-compliant models
• blocked by policies that restrict evaluation models

If none of Fusion's models are left, Fusion disappears from the model picker. In that case, the model policies set by your admins are the first place to look.

Local model routing is an experimental preview

On the same day, GitHub opened an opt-in experimental preview that adds models running on your own device to Fusion's pool. It works in Copilot CLI, the Copilot app and Visual Studio Code. On Fusion Standard, Copilot can now route requests to MAI Code 1.1 Flash or Qwen3.8 Flash Next on your own machine, but the routing decision still happens in the cloud, so sessions never run offline.

Local models Fusion can route to (GitHub Changelog)

Setup	Model	How you get it
Supported Windows devices	MAI Code 1.1 Flash	Download and configure through Windows ML
Copilot CLI on Windows, macOS or Linux	MAI Code 1.1 Flash or Qwen3.8 Flash Next	Configure the model yourself

Local models only come into play on Standard. GitHub also warns that they may have different safety protections from the cloud models Copilot provides, which is worth weighing before you turn the preview on for work code.

Start on Standard and switch up only for heavy tasks

Pro and lower plans only get Standard, and even on Pro+ and above GitHub positions Standard for everyday work and High or Max for complex bugs and multi-file changes. Because Standard uses one model per task, the draft-and-review patterns only come into play when you move up to High or Max.

Those higher levels can burn through AI credits faster, so the 30% discount that runs until November 30 is the cheapest window to test them. Local model routing, by contrast, only works on Standard.

If you used the HydraFusion preview, your old setting carried over to the matching level, which may be heavier than your daily work needs. Check the level shown in the model picker before you start.