GitHub Universe 2026: Copilot moves to the cloud, and sandboxes bill from November

What GitHub announced

GitHub posted 30 changelog entries on October 28, 2026, the day of its Universe keynote, most of them about Copilot. Its keynote post calls the whole an “agentic software delivery system.” Here is what changes for people who use Copilot, by GitHub’s own entries:

Announcement	Status on October 28	Plans
One cloud session across web, mobile, app, VS Code, CLI, Slack and Teams	Generally available, rolling out	Paid Copilot plans
Cloud sandboxes, now where those sessions run by default	Generally available; billed from November 2026	Paid Copilot plans
Copilot included with GitHub Enterprise, or $10 as Copilot Standalone	Through GitHub or Microsoft sales; self-serve “in stages”	GitHub Enterprise Cloud
Shared Copilot Automations, 24 new triggers, AI credit limits	Sharing rolling out; the rest available	Pro, Pro+, Max, Business, Enterprise
Projects in the Copilot app	Public preview	Not stated
Computer use in the Copilot app	Public preview, macOS and Windows	Not stated
Copilot code review Max	Public preview	Pro, Pro+, Max, Business, Enterprise
Copilot approvals on pull requests	Generally available, off by default	Pro, Pro+, Max, Business, Enterprise
Agent Merge in a pull request’s merge box	Available	Paid Copilot plans
Local models in Copilot Fusion	Experimental, opt-in	Not stated
GitHub Agentic Workflows	Generally available	GitHub Free, Team and Enterprise

The rest of this memo takes the ones that change what you pay or how you work, starting with where Copilot now runs.

Copilot moves to the cloud

“Copilot is now one connected experience in the cloud,” GitHub’s announcement says. A session you start on github.com, in GitHub Mobile, the Copilot app, VS Code, the Copilot CLI, or by mentioning @GitHub in Slack or Microsoft Teams is the same session on all of them, and every one shows the same session list. On your phone you can describe a task by voice or attach a screenshot, then finish it at your desk.

By default those sessions now run in a cloud sandbox, “an isolated Linux environment that GitHub creates for a Copilot session,” in the words of GitHub’s docs. Copilot clones your repositories into it and runs commands and tests there, so several sessions can run at once without your computer, and you don’t need to stay connected. Sandboxes have “no one-hour limit on total running session time,” and a stopped one keeps a snapshot, so you can resume it later from any client.

Some things stay as they were:

• The Copilot CLI still runs on your computer by default. copilot --cloud starts an interactive session in a sandbox instead.
• Some repositories still need Actions. Setup steps in copilot-setup-steps.yml, larger or self-hosted runners and Windows development environments work only when a session runs on GitHub Actions, which you choose when you start it.
• Copilot Free doesn’t get it. The new cloud experience is “available on paid Copilot plans only.” Free keeps Copilot Chat on github.com and GitHub Mobile.

It works the other way round, too. Register your own computer as an environment, with --environment-host in the Copilot CLI or code agent host --github-environment in VS Code, and you can start and steer the sessions on it from GitHub Mobile.

What cloud sandboxes cost from November

“Billing for cloud sandbox usage begins in November 2026,” GitHub says, without giving a day. It draws first on the compute allowance that comes with your GitHub plan, the same one that pays for Actions builds in private repositories. That allowance is now a dollar amount instead of a count of runner minutes. “It’s the same value,” GitHub says, and its billing docs list it by plan:

GitHub plan	Included compute a month	In standard Linux runner minutes
GitHub Free	$12	2,000
GitHub Pro	$18	3,000
GitHub Free for organizations	$12	2,000
GitHub Team	$18	3,000
GitHub Enterprise Cloud	$300	50,000

The allowance belongs to your GitHub account’s plan, not your Copilot plan. Someone paying $10 a month for Copilot Pro on a free GitHub account has $12 of compute a month, shared between builds and sandboxes, and GitHub says neither has priority.

Sandboxes are billed on three meters:

• Compute: $0.000024 a second while the session runs.
• Memory: $0.000003 a GiB-second of memory allocated to it, used or not.
• Storage: $0.005 a GiB-month for the snapshot of a stopped session.

GitHub’s own example is a session of 600 seconds with 4 GiB of memory, which costs $0.0216. By our arithmetic, an hour of that sandbox is about 13 cents, so $12 would cover roughly 90 hours of it if your builds used none. Past the allowance, usage is billed to your payment method; with none on file, it’s blocked. The repository’s owner pays: you for your own repositories, the organization for its own.

Sandbox compute is a separate bill from the model. A cloud session’s AI work still uses your Copilot plan’s AI credits, at one cent a credit, which the plans page puts at $15 a month on Pro, $70 on Pro+ and $200 on Max, part of it a “flex allotment” GitHub says “may change over time.”

Copilot included with GitHub Enterprise

The other change with a price on it is for companies. Since October 28, enterprises that buy through a GitHub or Microsoft account team can get “Copilot included with GitHub Enterprise for every licensed user,” says the changelog. GitHub calls it Copilot access, and its docs set it beside the seats it competes with:

Billing model	Seat charge, per user a month	Included AI credits
Copilot Business	$19	1,900 per user, pooled
Copilot Enterprise	$39	3,900 per user, pooled
Copilot access, with a GitHub Enterprise Cloud license	$0	None
Copilot access, as Copilot Standalone	$10	None

Copilot access has “the same features and models as Copilot Enterprise” but no included usage: AI usage “is metered from the first request.” When the budget runs out, code completion and next edit suggestions keep working, while Copilot Chat, code review and Copilot in the cloud wait for more budget. “An exhausted budget never removes access.”

The conditions:

• Sales only, for now. For self-serve enterprises it’s “currently unavailable,” and GitHub “will share dates in the coming weeks.” It isn’t offered on GitHub Free or Team, or on self-service Enterprise Cloud trials.
• One model per enterprise. You can’t mix Copilot access with Business or Enterprise seats, even in separate organizations.
• Seats still need assigning, licensed users included.
• Moving is voluntary. Business and Enterprise stay on sale, nobody is moved automatically, and a move takes effect on the first day of a month.

At list prices, a Business seat’s $19 buys 1,900 credits, worth $19 at a cent each, so the seat is mostly prepaid usage. Copilot access drops the prepayment for licensed users and bills what they use. GitHub’s docs walk through a four-step comparison against your current bill and say to confirm account-specific pricing with your account team. If your firewall allows only *.business.githubcopilot.com, add *.enterprise.githubcopilot.com before you move.

Automations, Projects and computer use

Copilot Automations are prompts with triggers and tools that Copilot runs in the cloud, “every day at 9am” or “when a pull request is opened.” Until now they were personal, running under your identity and your credits. Shared automations, rolling out on Pro, Pro+, Max, Business and Enterprise, belong to the repository: anyone with read access sees them, anyone with write access can edit them, their actions are attributed to Copilot, and their AI credits are billed to the repository’s owner. They keep running if their author leaves. In an organization, an administrator first has to enable “Allow use of Copilot cloud agent billed to the organization.”

Automations also gained 24 GitHub triggers, a REST API trigger for outside systems such as Sentry or PagerDuty, a way to run one automation after another, 20 new GitHub tools, tools from custom MCP servers, and AI credit limits per run, day, week or month. Automations without limits “continue to run as they do today,” so setting one is worth a minute.

In the Copilot app, Projects, in public preview, give work that spans several repositories one home, with its context, agents and results together. Computer use, in public preview on macOS and Windows, lets Copilot read and click through desktop apps, “including those without an API, command-line interface, or MCP integration.” It asks before it controls an app; turn it on under Settings > Computer Use. VS Code has an experimental version.

Code review Max and Agent Merge

Copilot code review has a third effort level, Max, in public preview. It “runs several independent review agents across models from different providers,” then reconciles what they find, and GitHub warns that it “takes longer and uses more AI credits” than Lite or Balanced. Copilot approvals are now generally available: once an admin turns them on, Copilot’s approval counts toward a repository’s required approvals, and new commits dismiss it, as they would a person’s.

Agent Merge now works from a pull request’s merge box. Choose Watch and fix, and Copilot fixes failing Actions workflows, merge conflicts or both in a cloud environment. It needs a paid Copilot plan.

Local models in Copilot Fusion

Copilot Fusion, GitHub’s router, which came within two points of Claude Code with Opus 5.5 on Artificial Analysis’s index, can now send requests to a model on your own computer, as an opt-in experiment. On supported Windows devices, Copilot downloads MAI Code 1.1 Flash through Windows ML. In the Copilot CLI on Windows, macOS or Linux, you can serve MAI Code 1.1 Flash or Qwen3.8 Flash Next yourself from a local OpenAI-compatible server. Choose Standard in Copilot Fusion, and it picks local or cloud for each request.

It doesn’t make Copilot work offline: “Copilot Fusion still makes routing decisions in the cloud.” GitHub also notes that local models “may have different safety protections” from its cloud models.

What to do before November

1. See where your sessions run. Cloud sessions started on github.com, in the Copilot app or in VS Code now run in a sandbox by default. If a task doesn’t need the cloud, run it locally in the CLI or the app.
2. Set a budget for sandboxes. GitHub’s “Bundled AI credits” budgets don’t cover them. Use a product-level or SKU-level budget, and turn on Stop usage when budget limit is reached if you’d rather it stop than bill.
3. Watch the shared pool. If your builds already use most of your Actions allowance, sandboxes will take you past it sooner.
4. Stop sessions you’re done with. Compute and memory stop when a sandbox stops; after that, only its snapshot is billed, until you delete it.
5. Put credit limits on automations, shared ones especially, since they bill the organization.
6. On an enterprise, estimate before you move. Run GitHub’s four steps on a full month of usage, counting what your pooled credits cover today, not only your overage.

What GitHub hasn’t said

• The day in November sandbox billing starts.
• How big a sandbox is. GitHub gives no vCPU or memory figures, and its two pages disagree on the unit: the billing page prices compute per “compute second,” the sandbox page per “vCPU second.” Until it says, what an hour costs is an estimate.
• When self-serve enterprises get Copilot access, beyond “the coming weeks.”
• Which plans get Projects and computer use in the Copilot app. Their changelog names none.
• What the Copilot SDK’s new name means for code. It’s becoming the Microsoft Copilot SDK, with “no immediate changes” required and no new package names yet.

If you use more than one agent

Copilot’s plans page lists delegating tasks to third-party agents, such as Claude and Codex, on Pro+ and Max, in preview. Enjoy takes a different route. It doesn’t run Copilot; it runs Claude Code, Codex, Grok Build, Antigravity, Muse Code and local models through OpenCode on your own computer, each on the plan you already have, and lets you switch agents in the middle of a conversation. Its scheduled workflows do for those agents what a scheduled automation does for Copilot, with one difference that matters here: they run on your computer, only while it’s awake with Enjoy open, rather than in a sandbox billed by the second. Enjoy is a monthly subscription, with the first seven days free (pricing).