GitHub’s outage report shows where Copilot model switching stops GitHub’s outage report shows where Copilot model switching stops

What failed in the October incident?

GitHub says an upstream credential became invalid. The provider then rejected Copilot requests to several GPT-family models, while other models remained available. Automatic selection redirected some traffic. Features tied to one model, including code scanning autofix, suffered more. Recovery required replacing the credential across environments. The October 28 incident was resolved before this report appeared.

What does automatic selection actually choose?

GitHub’s June 17 description of auto mode made request complexity and current availability inputs to model selection. The eligible pool depended on the subscription and configured policies, and could change over time. Users could inspect which model produced a response and choose a particular model for the next one. That explains the difference between a selection mechanism and a spare service. A router chooses among permitted destinations. Its usefulness depends on those destinations being available and suitable for the particular request.

How does Fusion change the choice?

GitHub launched Copilot Fusion on October 28, combining automatic selection with multi-model orchestration. Its three effort levels range from Standard to Max. A task can use one model, escalate a draft after a quality check, or send a draft to another model for review. Administrators can restrict Fusion’s choices independently of general model availability. An effort level describes how work is approached. The recovery question concerns which underlying service the work needs and whether another path can complete it.

What would a fallback test need to distinguish?

Consider a hypothetical coding service with two approved models and one shared credential. A routing test disables the first model and successfully reaches the second. A different test invalidates the credential needed by both. The same successful routing logic now has nowhere usable to send the request. Giving the second destination independent authentication changes that example again. These constructed cases illustrate a dependency problem without describing GitHub’s internal design. They illustrate why counting model names is insufficient: the number of independent working paths depends on the dependencies beneath those names. A successful answer and a successful recovery exercise answer different questions. The first shows that one request worked. The second identifies which failure the alternative could survive.

Why does the eligible model list keep mattering?

Scheduled retirement changes the list even without an outage. GitHub’s October 15 notice sets November 16 as the retirement date for Claude Opus 5 and GPT-5.6 Terra. It names newer alternatives, including Opus 5.5. Under default enablement, those alternatives become available to Business and Enterprise customers unless administrators have disabled them or the global default. That is a policy condition, not a guarantee inferred from the presence of a model picker. A planned migration and an emergency fallback both depend on the destinations actually allowed for the account.