Microsoft has begun routing a portion of user prompts in Excel and Word through its own MAI models rather than sending them to OpenAI or Anthropic, according to a Bloomberg report cited by TechCrunch on July 7, 2026. The change is described as a cost-saving measure, shifting inference load for at least some Office 365 features onto infrastructure Microsoft controls end to end.

Asked for comment, Microsoft told TechCrunch it had nothing further to share about the reported shift.

The move marks a reversal in public positioning. In September 2025, Microsoft advertised that large parts of Office 365 ran on models from both OpenAI and Anthropic, a dual-vendor arrangement that signaled comfort paying for frontier-model quality from external partners rather than betting solely on in-house silicon and weights. Ten months later, at least some of that traffic has been redirected internally.

The economics behind the pivot were flagged well before this week's report. Bloomberg reported on June 4, 2026 that Microsoft's AI chief said Anthropic's models were too expensive to run at Office scale. That comment followed Microsoft's Build conference, held in early June 2026, where the company launched seven new MAI models — a family that includes an agentic coder and a text-to-image generator, per TechCrunch and Microsoft's own announcement.

One of those seven, MAI-Thinking-1, is described by Microsoft as a midsize model, per Yahoo Finance's June 3, 2026 coverage of the Build launch. Midsize positioning matters here: it suggests Microsoft is not trying to match frontier-scale reasoning models on every axis, but rather to hit a cost-per-token and latency profile suited to high-volume, lower-complexity tasks — precisely the kind of routine cell-formula generation, document summarization, and formatting requests that make up a large share of Copilot's traffic inside Excel and Word. Euronews similarly reported the seven-model family as a direct competitive answer to OpenAI and Anthropic, rather than a research curiosity.

The groundwork for this predates Build by two years. Reuters reported in May 2024 that Microsoft was training a new in-house language model intended to be large enough to compete with Google's and OpenAI's offerings — well before the current cost-cutting framing took hold, and before Microsoft's own September 2025 marketing emphasized third-party model diversity. The trajectory from "training a competitive model" to "advertising a multi-vendor Copilot" to "quietly rerouting production traffic to in-house models for cost reasons" spans roughly two years and reflects a maturing MAI stack rather than a sudden reversal.

The commercial logic is straightforward for anyone who has tracked inference economics at hyperscaler volume. Microsoft pays OpenAI and Anthropic per token for every Copilot request routed to their models, on top of whatever revenue-sharing or investment arrangements already exist between the companies. Every prompt shifted to a self-hosted MAI model run on Microsoft's own Azure capacity removes a variable cost paid to a partner and converts it into a largely fixed infrastructure cost the company already controls. At Office 365's user base, even single-digit-percentage traffic migration compounds into meaningful savings.

What Microsoft has not disclosed is more consequential than what it has. The company declined to specify what share of prompts now route to MAI models, which specific Copilot features are affected, or whether output quality benchmarks were part of the routing decision versus cost alone. Absent that detail, it is not possible to say whether users are getting a comparable experience at lower cost to Microsoft, or a somewhat degraded one.

Worth flagging: the OpenAI and Anthropic partnerships were never purely technical arrangements. Microsoft holds a substantial equity stake in OpenAI, and any diversion of production workload away from that dependency has implications for the partnership's economics beyond simple vendor competition. A model-routing decision inside Excel is, in that sense, also a signal about how Microsoft weighs the value of external frontier-model access against the cost of building competitive capability in-house — a calculation every large AI consumer, not just Microsoft, is currently running.

The broader pattern is not unique to Microsoft. Enterprises with the scale to justify it have increasingly moved toward hybrid model strategies: expensive frontier models reserved for complex reasoning tasks, cheaper in-house or open-weight models handling routine volume. Microsoft's position is distinctive mainly because of scale — few companies have both the Office 365 user base to generate the cost pressure and the MAI research investment to produce a credible substitute. Whether that substitute holds up under the kind of scrutiny frontier models routinely receive is the open question analysts and enterprise customers will now be watching for in coming quarters.