News · 2026-10-29

GitHub connects Copilot cloud work across devices and prices model orchestration

GitHub made connected Copilot cloud sessions generally available on paid plans on October 28, letting developers continue the same work across supported devices and interfaces. The wider launch also includes model orchestration through Fusion, while Projects and desktop computer use remain in public preview. The change makes a persistent task, with its own state and costs, the center of the coding workflow.

Key facts

• Connected cloud sessions are generally available on paid Copilot plans, subject to organization policy.
• Projects and desktop computer use have separate public-preview status.
• GitHub’s Fusion launch advertises a temporary 30% discount on paid plans, with every selected model billed.
• Primary source: GitHub’s October 28 connected-cloud changelog.

GitHub’s invitation to “Work from anywhere” is more specific than remote access to a chat window. Its changelog says a developer can start a cloud task through github.com, GitHub Mobile, the Copilot app, the command line, VS Code, Slack, or Microsoft Teams and continue that same session elsewhere. The agent’s work runs in a GitHub-managed cloud sandbox by default, with saved snapshots supporting resumption.

Think of the difference between carrying a notebook between desks and leaving a staffed workbench running in a workshop. A chat history carries what was said. A persistent cloud session can also preserve the working environment and the state needed to continue a job. That does not guarantee a successful result, but it changes what a device switch or a closed laptop means for an ongoing task.

Mobile is part of that design rather than merely a place to read notifications. GitHub says users can dictate a task or attach a file or image, and can steer a local machine registered as an environment. Our lesson on agent harnesses explains why these surrounding execution and control systems can matter as much as the underlying model.

The Copilot app release adds a different layer. Projects coordinate work across repositories in managed parallel sessions and keep progress and decisions together. GitHub labels Projects public preview. Computer use, which can read and control desktop applications, is also a public preview on macOS and Windows and asks for user approval before controlling an application. Neither feature should inherit the generally available status of connected cloud sessions merely because the announcements appeared together.

A third layer is Copilot Fusion. The user selects an effort level, and Fusion can choose one model or orchestrate several through a cascade or critique process. This reduces the need to choose every model manually. It also means one visible task may involve more than one model’s work behind the interface.

The billing consequence is explicit: each model Fusion uses is charged at its standard rate, with the limited-time launch discount on paid plans. This is not a promise of a flat price per completed task. Separately, cloud sandboxes draw on a shared dollar-based Actions compute entitlement. Model work and the computer on which that work executes are different cost categories, even when the product makes them feel like one continuous session.

GitHub reports favorable cost and time results for Fusion on Artificial Analysis’s coding-agent evaluation. Those are GitHub’s attributed benchmark claims, not a guarantee that a particular repository task will cost less. A critique pass can improve a result while adding calls; extra parallel work can reduce elapsed time while raising total compute. The useful comparison is completed, acceptable work at a known budget.

The broader GitHub Universe announcement frames the launch as an evolution toward orchestrating agents. The strongest counterargument is operational: continuity can make it easier to leave work running without noticing accumulated cost or a stalled plan. More agents also create handoffs and permission questions, as our multi-agent systems lesson explains.

The launch also makes resumption a product property worth testing. A saved snapshot should be judged by whether the task can actually continue with the needed files and state, rather than whether the interface still displays a conversation. GitHub’s announcement establishes the feature, while task-specific reliability remains something users must assess.

The release therefore matters beyond a longer feature list. It makes persistence, review, and accounting part of the same developer surface. Its practical limits remain feature-specific availability, organization controls, and the difference between an agent producing a proposed change and a human deciding that change should ship. Developers should assess the connected session, the preview coordination features, and Fusion’s billing as related but distinct parts of the launch.

Primary source, verified: read the paper →

Ground Truth. (2026, October 29). GitHub connects Copilot cloud work across devices and prices model orchestration. Ground Truth. https://groundtruth.day/news/github-copilot-connected-cloud-fusion.html

@misc{groundtruth:github-copilot-connected-cloud-fusion,
  title  = {GitHub connects Copilot cloud work across devices and prices model orchestration},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/github-copilot-connected-cloud-fusion.html}
}

Topics: github · coding-agents · developer-tools · cloud · model-routing