GitHub Turned Copilot Into a Project Manager, and the Bill Quietly Halved

At GitHub UNIVERSE'26 on October 28, GitHub released five Copilot-related things in one go. The list is not long, but together they point to the same thing: Copilot is no longer just the tool in the sidebar that helps you complete code.

The five features are:

• Copilot Fusion: When a single model is not enough, it coordinates multiple models and workflows.
• Copilot Projects: See where work stands and direct what Copilot should do next in one place.
• Copilot Security Extension: Bring code the agent just wrote and repository code-scanning alerts onto the same canvas for review.
• Copilot dreaming: Suggest improvements to personal memory, repository instructions, agent skills, and custom code review instructions.
• Copilot Automations: Define a task once, then run it on demand, on a schedule, or in response to events in GitHub and connected services.

Then there are the Fusion numbers. GitHub cited Artificial Analysis's Coding Agent Index: Copilot Fusion Max is within two points of Claude Code with Opus 5.5 Max, and its API cost per task is about half of the latter. There is a 30% discount before November 30.

One more thing needs to be added here. In its own reply, GitHub states that this evaluation was commissioned and funded by GitHub, and that the full results and methodology are on Artificial Analysis's page. There is nothing particularly questionable about the numbers themselves, but they need an independent rerun. Someone in the comments has already pointed out that recalculating using cost per solved task would be more convincing. If the conclusion holds up, at least some Claude Code users will come try it.

A few details worth watching.

Routing is opaque. Some users asked whether Fusion can show which model wrote which part of a change, and when routing stays with a single model versus switching to review. No answer is visible yet. This determines whether it saves money on messy tasks or only on clean benchmarks. Another person mentioned long batch runs lasting many hours—those are the scenarios that actually generate the bill.

Fusion can now route to local models, in an experimental phase, but according to the changelog, the routing call itself still happens in the cloud. Note that.

dreaming is the quietest of the five. Between sessions, the agent proposes changes to its own memory, skills, and instructions—meaning it can get stronger without changing models. Conversely, one bad edit can quietly accumulate. One user put it very accurately: the diffs to these files should go into review even more than code diffs do.

Projects, by contrast, was mentioned the most in the comments. One developer said the most troublesome part of using an agent is not writing code, but not knowing where five things stand when they are running at the same time. GitHub's solution is a dashboard. Someone joked that the construction crew has finally hired a foreman with a clipboard. There was an even sharper version: that ticket marked in progress since March finally has its own kanban board.

Automations and Projects are a matched pair: one defines how a task runs, and the other shows how far it has gotten.

Copilot is becoming an entry point. This direction is not new, but this time it has put entry point, orchestration, memory, auditing, and automation all on the table. The unanswered questions are still the same few: who decides which model is used, can every step be traced, and where does the money go on long tasks. And that evaluation—when will someone run it independently?