Microsoft routes Copilot to its own models — the largest buyer stops being a buyer
Microsoft has moved production Copilot traffic onto its in-house MAI models, with MAI-Code-1-Flash becoming the default GitHub Copilot model. The MAI family unveiled at Build 2026 is described as trained entirely in-house with no distillation from third-party labs. The significance is structural: the industry's largest distributor of another lab's models is reducing that dependency.
What changed
Microsoft's MAI family, unveiled at Build 2026, is described by the company as seven cloud-tier models trained entirely in-house with no distillation from third-party labs. Through August, Microsoft moved production traffic onto them: MAI-Code-1-Flash became the default GitHub Copilot model, and Microsoft has described routing large volumes of Excel and Outlook prompts to its own models rather than third-party APIs.
Why this is a structural story rather than a product story
Microsoft has been the single largest channel through which another lab's models reached enterprises. A shift to first-party models at the default tier changes what "AI in Microsoft products" means, without most users noticing anything.
For a buyer, the practical consequence is that the model behind a feature you depend on can change without the feature changing name. If you validated a workflow against particular model behaviour, that validation was against something the vendor can substitute.
Read this alongside the Cursor decision
Two stories in the same week, pointing the same direction. In one, a lab withdrew model access from a customer over corporate ownership. In the other, the largest distributor reduced its dependence on an outside lab.
The common thread is that the model layer is not a stable commodity. It is contested, and it is moving. Any industrial AI plan with a multi-year horizon should assume the specific model behind it will change at least once, and should be architected so that changing it is a configuration exercise rather than a revalidation project.
That is not a reason to delay adoption. It is a reason to keep the model replaceable.
Sources
- Microsoft AI — Building a hill-climbing machine: launching seven new MAI models
- Microsoft AI — Microsoft Build 2026: MAI keynote transcript
Verified 2 September 2026. Benchmark and routing claims in this entry are Microsoft's own published statements, not independent measurements.