September 14, 2026: GitHub recovers from a critical service incident

Today is September 14, 2026, and GitHub’s biggest fresh signal is operational. The status page now says All Systems Operational after a critical incident on September 13 that affected API Requests, Issues, Pull Requests, Actions, and Pages.

GitHub says the incident was resolved at 10:44 UTC. In its disclosure, the company points to increased database replication delays on collab, which then caused authorization endpoint errors and follow-on errors across the system. That matters because it is not just one visible page or one API route failing in isolation. It is a reminder that a shared backend issue can spread into collaboration, automation, and publishing features at the same time.

Here is the plain-language mechanic. When a database replica falls behind, it may answer with stale state or missing state. If another service depends on that state to check identity, permissions, or object existence, requests can start failing. Then clients retry, traffic rises, and the original lag can turn into wider errors. That is the failure pattern GitHub described here, and it is the main reliability lesson from the weekend event.

The important thing for users is that the service is now marked healthy again. If you saw odd authentication behavior, broken issue or pull request actions, API errors, or Pages problems during the incident window, GitHub has now moved the platform back to green. But resolution is not the same thing as explanation. The next artifact that matters is still missing: a fuller root-cause writeup that names the trigger, the duration, and the hardening steps. Until that appears, the incident should be treated as resolved but not fully explained.

That distinction matters for operators. A status flip tells you the immediate outage is over, but it does not yet tell you whether the underlying pattern was a one-off, a capacity problem, a data consistency issue, or something else in the service chain. For teams that rely on GitHub for code hosting, reviews, automation, and publishing, the useful question is not just whether the page is green. It is whether the failure mode could recur in a similar shape.

On the product side, there is no fresh launch burst in the public changelog that I could confirm today. The visible changelog archive still tops out at September 11 entries, including the VS Code Agents usage-metrics update and Copilot code-review updates. So the week’s center of gravity is still around operational reliability and Copilot controls rather than a major new feature push.

That Copilot control surface remains worth watching. GitHub’s supported-models page still lists MAI-Code-1-Flash as retired on September 10, and it says Claude Sonnet 4.6 remains available to annual Copilot Pro and Copilot Pro+ subscribers. That is a small but material reminder that the model menu inside Copilot is not static. Providers come and go, and entitlement rules matter as much as the model names themselves.

For users, that means the exact experience can differ by plan and by surface. A model being listed in documentation does not guarantee it appears everywhere in the same way, and a retirement notice does not tell you how fast every client or admin setting updates. The unresolved check here is the exact day-of-change behavior across github.com, Copilot CLI, VS Code, and enterprise-managed settings after the MAI retirement. That still needs direct confirmation from screenshots or updated docs.

Put differently, GitHub is managing two kinds of trust at once. One is platform trust, which depends on the service being available when people need to review code, open issues, run actions, or publish pages. The other is product trust, which depends on Copilot behaving in a way admins and subscribers can understand. Those are different surfaces, but they both affect whether customers feel in control.

The broader pattern is familiar. Reliability incidents and model-menu changes may sound unrelated, but both shape the developer experience. One affects whether the service is there when you need it. The other affects whether the AI features are available in the way you expect. In both cases, the key question is whether GitHub’s documentation and behavior stay aligned with what customers actually see.

So the scorecard for today is short. GitHub is back up. The incident was in core collaboration plumbing, not a cosmetic glitch. The likely failure mode involved replication lag cascading into auth and other errors. And Copilot’s model lineup is still being actively managed, with retirement and entitlement details that admins should keep an eye on.

What to watch next is concrete. First, whether GitHub publishes a root-cause or postmortem that explains the September 13 incident in more detail. Second, whether the Copilot model-picker behavior is documented clearly across web, desktop, CLI, and enterprise settings after the MAI-Code-1-Flash retirement. Third, whether the changelog stays quiet or turns back up with another Copilot or admin release before the end of the week.