Microsoft Routes Office Tasks to Its Own AI to Cut Costs

Microsoft has started sending some user requests in Excel and Word through its own AI models instead of to OpenAI or Anthropic, according to Bloomberg reporting cited by TechCrunch on July 7, 2026. The shift appears designed to save money by handling at least some AI work on infrastructure Microsoft owns and controls, rather than paying per request to external partners.

Microsoft declined to comment beyond confirming the change to TechCrunch, offering no detail on how much traffic has moved or which specific features are affected.

The Shift Reverses Recent Strategy

Just ten months ago, in September 2025, Microsoft was publicly highlighting that Office 365 relied on models from both OpenAI and Anthropic. That dual-vendor setup suggested the company was comfortable paying for high-quality external models rather than betting everything on its own in-house technology. The latest move signals a recalibration.

The economics that drove this decision surfaced in public months earlier. On June 4, 2026, Bloomberg reported that Microsoft's AI chief said Anthropic's models had become too expensive to operate at the scale of Office 365 — where hundreds of millions of users generate requests every day. That comment came right after Microsoft's Build conference in early June, where the company unveiled seven new in-house AI models, including one called MAI-Thinking-1, described as a midsize model.

Understanding the Strategic Fit

Midsize positioning is meaningful. Microsoft is not trying to build models that outperform the most advanced external options on every measure. Instead, the company is optimizing for cost and speed on the types of tasks that Office users generate in volume: writing formulas in Excel, summarizing documents, adjusting formatting. Those routine requests don't require the kind of advanced reasoning power that frontier models provide — and they don't need to.

This strategy didn't emerge overnight. Reuters reported in May 2024 that Microsoft was already training its own competitive language model, well before the company's September 2025 marketing pushed the multi-vendor Copilot story. The path from "building a competitive model" to "advertising OpenAI and Anthropic diversity" to "quietly redirecting routine traffic to our own models" took about two years and reflects a step-by-step maturation of Microsoft's AI capability rather than a sudden reversal.

The Math Behind the Decision

The financial logic is straightforward for large-scale operations like Office 365. When Microsoft routes a request to OpenAI or Anthropic, the company pays a fee per token — the small units of text that AI models process. Each prompt sent instead to a Microsoft-owned model removes that variable cost and converts it to a fixed infrastructure expense the company already covers. At Office 365's user base, even routing a small percentage of traffic this way compounds into millions in annual savings.

What Remains Unknown

Microsoft has not disclosed what percentage of prompts now use in-house models, which Copilot features are affected, or whether the decision was based purely on cost or also involved comparing output quality. That omission matters. It leaves open whether users are getting equivalent results for Microsoft's cost savings, or a somewhat lower quality experience while Microsoft pockets the difference.

The Broader Stakes

The partnerships with OpenAI and Anthropic were never just technical arrangements. Microsoft holds a significant ownership stake in OpenAI. Any substantial shift of workload away from OpenAI's models carries meaning beyond ordinary vendor competition — it reflects how Microsoft is internally weighing the value of access to the best external AI capabilities against the cost of building strong alternatives in-house. Every large company making heavy use of AI is wrestling with this same calculation right now.

The larger pattern extends beyond Microsoft. Organizations with sufficient scale are increasingly adopting hybrid approaches: keeping the most expensive frontier models for complex reasoning tasks, using cheaper in-house or openly available models for high-volume routine work. Microsoft's position stands out mainly because of sheer scale — few companies operate both the global user base that creates cost pressure and the research resources to build credible alternatives. Whether Microsoft's in-house models can perform consistently enough to justify this shift at enterprise scale is what analysts and customers will be watching in the quarters ahead.