GitHub says Copilot cloud sandbox billing begins November 23, 2026. Its connected Copilot cloud experience uses cloud sandboxes for agent sessions, and usage can draw from the same dollar-based GitHub Actions compute entitlement used for builds. That makes this more than a purchasing change: it is a testable metering and governance change for teams that run both CI and AI-assisted work.
What changes on November 23
GitHub documents that cloud sandbox usage begins billing on November 23 and draws from included Actions compute. The shared entitlement covers standard GitHub-hosted runners, cloud sandboxes, and—on Team and Enterprise Cloud—larger hosted runners. Each product retains its own published rate; higher-rate compute consumes the shared entitlement faster. Usage beyond the entitlement follows the account’s existing rates, budgets, and payment settings.
This does not mean a sandbox minute and a runner minute are interchangeable. It means QA should validate the account-level outcome using the applicable published rates and billing views.
Copilot cloud sandbox billing QA checklist
- Establish a baseline: record current Actions entitlement, budget rules, payment settings, and usage before enabling a controlled sandbox workload.
- Use a disposable test repository: run a small, bounded Copilot cloud task and capture its start/end time, environment choice, commands, and final pull request or issue outcome.
- Separate workloads: execute a known CI workflow before and after the sandbox test. Confirm that the reporting and accounting views distinguish the product activity as documented while the shared-pool balance changes as expected.
- Test the boundary: use a low, non-production budget threshold or a dedicated test organization where permitted. Verify warning, blocking, and overage behavior without exhausting a live team’s allowance.
- Check access controls: confirm that the Copilot in the cloud policy, repository exclusions, agent permissions, and MCP allowlists still apply. Cost controls are not a substitute for authorization controls.
- Reconcile evidence: compare the user-visible session record, Actions usage/billing evidence, and repository change. A successful agent response alone is not proof that the workload was correctly attributed.
Why this matters for QA engineers
AI-agent testing increasingly spans application behavior, permissions, and operational controls. A shared compute pool can create a cross-system failure mode: an unexpected agent workload may affect CI capacity or spending visibility even when the code change itself is harmless. Treat billing, policy enforcement, and audit evidence as acceptance criteria. The goal is not to predict a bill from one test, but to prove that the organization can explain, constrain, and reproduce the path from a cloud session to consumption.
A small acceptance test to run
Given a test organization with a documented compute balance
When one approved Copilot cloud sandbox task runs in a disposable repository
And one known GitHub Actions workflow runs
Then usage evidence identifies the relevant product activity
And the shared entitlement changes consistently with published rates
And policy and budget controls produce the expected result
And the repository outcome is independently verified in clean CI
Common mistakes
- Assuming every cloud-session action is billable in the same way as a standard runner.
- Testing in a production organization with real budget limits and no rollback plan.
- Using only a successful pull request as evidence of correct metering.
- Ignoring organization policies, repository exclusions, SSO, or MCP restrictions while testing cost behavior.
Bottom line
Copilot cloud sandbox billing is a timely reason to add financial and operational assertions to agent QA. Start with a small reproducible workload, preserve before-and-after evidence, and validate the shared compute pool alongside normal CI. That gives teams a defensible rollout signal before agent use becomes routine.
Sources
- GitHub Changelog: Work from anywhere with a new, connected Copilot cloud experience (updated November 19, 2026)
