Microsoft Begins Replacing OpenAI and Anthropic with Its Own AI in Office

In recent hours, Microsoft has started serving a portion of prompts for some Office applications with its own internally developed AI models (MAI), instead of those from OpenAI and Anthropic. The shift pertains to tasks where cost and data residency favor in-house production; this was revealed by Bloomberg.

OpenAI and Anthropic models continue to handle the majority of the traffic within Copilot, and the MAIs are integrated only where it makes financial sense. Bloomberg reports that tens of thousands of prompts per week have already been completed by the internal models in Excel and Outlook, still a small fraction of the total.

The direction had already been set: in June, the head of the company's models, Mustafa Suleyman, had openly stated the desire to reduce spending on Anthropic by shifting workloads onto the MAIs.

From Dependency Built Over Years to Proprietary Models

For years, the original contract with OpenAI prevented Microsoft from pursuing cutting-edge AI independently. The 2025 renegotiation ended the exclusivity, granting the company a license to OpenAI's technology until 2032 and, along with it, the freedom to build competing models.

This has led to a three-pronged strategy: a significant stake in OpenAI, Anthropic's Claude integrated into Copilot, and now the in-house MAIs. It's the logic that Satya Nadella would have warned against reducing Microsoft "to the next IBM" if it relied too much on a single partner: owning the model, rather than merely renting it, is the way to avoid that fate.

Distribution Matters More Than Benchmarking

Microsoft does not need the MAIs to dominate the rankings. The leverage is distribution: hundreds of millions of Office and Teams installations, where shifting even a portion of traffic to internal models translates into tangible numbers.

At June's Build conference, the company presented seven MAI models, from the first reasoning system to image generation, voice, transcription, and an agentic coder. It claims that an MAI tailored to Excel matches a GPT-5 generation model with efficiency up to ten times higher. On the coding front, according to Bloomberg, an MAI model reaches the capabilities of Opus 4.6 at a lower cost.

This move aligns with a trend of greater frugality in the industry, with several major companies in recent months trying to curtail AI spending after last year's surge. The expenses for third-party models are not trivial, and every function brought in-house diminishes that cost. However, the lingering question that Microsoft has yet to resolve remains: who will truly pay for all these functions? In-house models are more economical and help margins regardless of the answer.