The Black Cloud

AI - Copilot Fusion Max and GitHub Universe 2026, a bumpy start but a good result 🤣

I got automatically switched to Copilot Fusion in VSCode, configured it to Max, and initially it broke quite a lot of my workflow 🤣 Not exactly the introduction I was hoping for.

After a few tweaks, though, it worked really well. So this is not a story about everything being perfect from the first prompt, but it has certainly made me interested in what GitHub is doing here.

Copilot Fusion

Fusion is not another foundation model. It is a system that decides which models to use and how they should work together.

I covered Copilot’s earlier Auto model selection tiers before. Fusion brings that model selection together with Project HydraFusion’s orchestration: depending on the task and effort level, it can go beyond choosing one model and arrange drafting, checking and revision.

Official announcement

Route and orchestrate models with GitHub Copilot Fusion

What is new?

The three selectable effort levels are Standard, High and Max. Standard chooses one model for everyday work. High and Max can add orchestration for harder tasks, with Max prioritising the strongest capabilities rather than the lowest possible cost.

GitHub describes three ways to execute the work:

The model pool can change over time. Fusion also considers availability and prompt caching, so it is not simply switching models for the sake of it.

The release notes explicitly describe automatic migration from earlier Auto and HydraFusion selections for research-preview participants. That helps explain why the model picker can change without you deliberately choosing a new model.

NOTE

Fusion is generally available. Standard is available on all Copilot plans; High and Max require Copilot Pro+, Copilot Max, Copilot Business or Copilot Enterprise, subject to model policies.

To try it in VSCode, update the client, use a Copilot harness session, select Copilot Fusion in the model picker and choose an effort level. After a response finishes, hover over the effort-level label below it to see which models were used.

The performance and cost claims

In the @ArtificialAnlys Coding Agent Index, Copilot Fusion Max scored within two points of Claude Code with Opus 5.5 Max at about half the API cost per task.

The launch result is 60 for Fusion Max versus 62 for Claude Code with Opus 5.5 Max. Artificial Analysis tested Fusion Max in Copilot CLI and reported $5.59 per task versus $11.42 for that Claude Code configuration. GitHub also reports about 40% less time per task in its launch comparison.

Those are benchmark results, not measurements from my own workflow. They are impressive, but they do not mean every task is faster or every Copilot bill is halved.

The launch posts also use an up-to-77% lower-cost claim. For Max, I would keep the clearly stated roughly-half-cost comparison separate rather than assume that larger saving applies to my work.

NOTE

Through 30 November 2026, users on paid Copilot plans receive a 30% discount on model costs when using Fusion. This is a temporary billing promotion, separate from the benchmark API-cost comparison.

Normally, Fusion bills each underlying model at its standard rate, with no separate Fusion charge. High and Max can make additional model calls, so a demanding task can still consume more AI credits than it would on Standard.

Encouraging reactions

Scott Hanselman has had longer with it:

Half the cost for 2 points. I’ve been using it for weeks, it’s very good

Vlad raises the practical question behind the model orchestration: can it retain enough caching efficiency to make the extra coordination worthwhile? GitHub’s documentation says routing takes cache boundaries into account. The cost of the completed task matters more than the number of impressive model names involved.

Chubby also discusses Gemini 4 Carbon in that second post. I covered Carbon separately, including the caveat that Argon still had the lower hallucination rate in the cited Artificial Analysis evaluation.

Not everyone is convinced

Copilot Fusion makes no sense. The value is it’s slightly cheaper. Who cares, slightly cheaper only slightly matters. The open weight models are going to annilate cost. This is such a naive short term play. In 1-2 months Copilot Fusion will be irrelevant.

Very interesting take 🤣 There is a fair question here about how durable the saving will be as model prices change.

But open-weight models and orchestration are not opposites. The current Fusion documentation already describes preview support for routing to supported local models on Standard, not Max. The model pool is not fixed, so I would not assume the idea becomes irrelevant simply because cheaper models arrive.

For me, the test is whether it produces useful work and follows my instructions. My first experience was rough, but after those tweaks it worked really well. That matters more to me than winning an argument about a leaderboard 😊

Other GitHub Universe updates worth a look

The official Universe announcement and event recap cover much more than Fusion. These are the parts that stood out as relevant to the wider workflow.

Projects, automations and a shared harness

Continuing work across devices

The VSCode Universe update lets you follow the same cloud agent session across VSCode, Copilot CLI, the Copilot app, GitHub Mobile and the web. Sessions started through Slack or Teams can also be continued in those clients.

There is an important distinction here: cloud work can continue with your laptop closed; agents on your own or remote machines still need their host machine and Agent Host running.

The Copilot harness, built on the Microsoft Copilot SDK, is also now the default for new VSCode sessions. There is more changing here than just the name in the model picker.

Keeping the work reviewable

The wider Universe recap includes some useful changes around the work itself:

GitHub also outlined its infrastructure expansion and ongoing move to a multi-region Azure deployment. More agent activity needs a reliable platform underneath it, not just another way to generate code.

GitHub Universe keynotes: Day 1 and Day 2

Both keynote broadcasts are here, along with GitHub’s written Day 1 recap.

Day 1

Day 2

<< Previous Post

|

Next Post >>

#Blog #AI #GitHub #GitHub Copilot #Copilot Fusion #Copilot Fusion Max #GitHub Universe #VSCode