Microsoft uses its own MAI models in Office to cut AI costs

Microsoft has begun routing some user requests in Excel and Word to its own MAI model family instead of relying exclusively on OpenAI and Anthropic APIs. The company launched seven MAI model variants earlier in 2026 and is now deploying them in production for a portion of Copilot 365 traffic. The shift is explicitly cost-driven: third-party AI inference at enterprise scale has become expensive, and Microsoft is not alone in rethinking the calculus.

Amazon, Uber, Meta, and Accenture have made similar moves in recent months, substituting proprietary or open-weight models for third-party APIs in high-volume workloads where cost per token accumulates quickly.

For product managers working inside companies that have built features on OpenAI or Anthropic endpoints, the pattern illustrates a decision point that arrives as AI products mature. The early phase of AI product development often justifies premium API costs because iteration speed and model quality matter more than margins. As a product matures and usage scales, the cost structure that made sense at launch can erode unit economics significantly.

The more durable lesson is about dependency planning. Teams that architect features with the API provider abstracted behind a thin interface — rather than hardcoding model-specific behavior throughout the stack — preserve optionality. When the cost case for switching arrives, the migration remains tractable. Microsoft’s move confirms that the cost case arrives eventually for almost every AI product operating at scale.