August 31, 2026: GitHub’s late-August Copilot checklist

Good morning. It is August 31, 2026, and GitHub’s freshest operating signal is small but real: a minor Copilot model-provider incident briefly pushed errors for some OpenAI-backed models before service returned to normal within 43 minutes.

That matters less because of the duration, and more because it shows how much of Copilot still depends on a moving layer of model providers. When that layer hiccups, users may see a degraded agent or chat experience even if the rest of GitHub is healthy. GitHub’s status disclosure does not say anything broader than the incident window and the affected model group, so the right reading is simple: short outage, narrow scope, resolved.

The bigger story is the late-August policy and product cleanup around Copilot and adjacent surfaces. Today is the final export day for GitHub Spark. GitHub stopped allowing new users or new apps on August 4, and told existing users to export their apps by August 31. Deployed apps are still expected to keep working after retirement. In plain terms, that means GitHub is closing the creation path but not immediately pulling the plug on what already exists.

At the same time, GitHub is still rolling out its global model policy for Copilot Business and Copilot Enterprise through September 1. The practical effect is that open-weight models and models covered by GitHub’s data-retention gap are disabled by default unless an admin already chose otherwise. For enterprise buyers, the key question is not abstract model quality. It is which models remain selectable by default, and whether an admin has to re-approve anything before users notice a change.

Billing is also changing in a way that should be visible to buyers, even though list prices are not changing. GitHub says Copilot Business and Enterprise signups reopen on September 1 for credit card and PayPal buyers, but new seat access will require upfront payment. Existing card and PayPal customers will begin upfront charges on October 1. That is a workflow change more than a price change, but it still affects procurement timing, budget planning, and how quickly teams can add seats.

There is also a useful mechanic underneath all this. GitHub bills Copilot usage in AI credits, and one credit equals one cent. Those credits pool across the billing entity, so a heavy agent session can burn through the shared bucket faster than headcount alone would suggest. That is why the billing changes, budgets, and model controls matter together. They are not separate admin toggles. They jointly determine how quickly usage turns into cost.

GitHub has also been widening the control surface inside Copilot itself. In the latest Visual Studio update, GitHub added org-level custom agents, usage visibility, thinking-effort controls, richer model controls, and Git agent review of commits and uncommitted changes. And in the weekly Copilot releases, GitHub added shared team sessions in Slack and Teams, a general-availability Customize tab for MCP servers, plugins, skills, and canvases, faster CLI sessions, and more model and session controls in VS Code and Visual Studio.

The direction is clear. GitHub is pushing Copilot from a single chat box toward a managed, multi-surface workflow. That creates more places where admins can shape behavior, but also more places where defaults can shift under users if they do not watch the release notes closely.

Two other changes are worth keeping on the board. Starting October 1, Actions retention will cover checks, workflow runs, and statuses, using the same retention setting that already applies to artifacts and logs. GitHub says the default remains 90 days. In plain language, that means the same cleanup policy will govern more of Actions history, which can reduce old metadata piling up and may change how long teams can inspect past runs.

And GitHub Apps can now access enterprise billing data through installation tokens with the enterprise billing permission. That is a small but important infrastructure change because it lets finance and BI automation avoid depending on one person’s token. For enterprises, that is usually what turns a brittle one-off report into a durable internal workflow.

The backdrop to all of this is scale. In Microsoft’s July 29 earnings call, the company said GitHub Copilot now has 50 million users, Copilot revenue accelerated more than 60 percent quarter over quarter, and one in three GitHub pull requests now involves an agent. Microsoft did not break out GitHub revenue or margins, but those figures explain why even small policy, billing, and model changes have outsized operational importance.

So the day’s headline is not a major outage or a new flagship launch. It is a reminder that GitHub is in the middle of a dense rollout calendar: Spark export today, model-policy enforcement into September 1, billing changes on September 1 and October 1, Actions retention on October 1, and a Copilot review default change on September 28.

What to watch next is straightforward. First, whether GitHub publishes day-of-change documentation showing exactly which Copilot models remain selectable after September 1. Second, whether a billing or service postmortem appears for the August 27 incident thread. Third, whether GitHub confirms the September 28 Copilot code review default shift and the opt-out path for admins. And fourth, whether Spark export retirement stays clean, with no report of broken migrations or failed app access after today.