Microsoft Starts Swapping OpenAI and Anthropic for Its Own MAI Models to Cut AI Costs

Microsoft has started routing some prompts in Excel, Word and Outlook to its own in-house MAI models instead of OpenAI or Anthropic, TechCrunch reported, joining a broader industry move to cut AI costs. Microsoft AI chief Mustafa Suleyman told Bloomberg the company wants to eventually stop paying Anthropic altogether, calling its models "extremely expensive."

Definition: MAI is Microsoft's own family of AI models — covering reasoning, coding, image generation and voice — built to reduce reliance on OpenAI and Anthropic in Microsoft's own products.

Example: Some Excel and Word prompts that used to go to OpenAI or Anthropic are now being answered by Microsoft's in-house MAI models instead.

Key takeaway: This is a cost-driven shift, not a full break — Microsoft is still using third-party models alongside its own, but is deliberately growing the in-house share.

Business impact: Teams building on Microsoft 365 Copilot or Azure AI should expect more of the underlying model mix to quietly shift toward Microsoft's own models over time, which can change output style and cost even where the API surface stays the same.

What did Microsoft actually change?

Microsoft has begun using its own MAI models to answer a share of user prompts inside Excel and Word, tasks previously handled by OpenAI and Anthropic models. The shift is partial, not a full replacement: Microsoft is still relying on third-party models for other workloads, and the reporting frames this as an early, deliberate rebalancing rather than a finished migration. Microsoft did not offer additional comment beyond what's publicly available when asked about specifics.

What are the MAI models replacing OpenAI and Anthropic?

Microsoft launched seven MAI models at its Build conference: MAI-Thinking-1 for complex reasoning, coding and multi-step problems; MAI-Code-1 for software development, tuned for GitHub and available in Copilot and VS Code; MAI-Image-2.5 for text-to-image and image-to-image generation, live in PowerPoint and rolling out to OneDrive; plus additional models for voice and transcription. Microsoft has said the same models are expanding toward GitHub Copilot and Teams beyond Excel, Word and Outlook.

Why is Microsoft doing this — and why now?

Cost is the stated reason. Suleyman told Bloomberg that "Anthropic is extremely expensive and I think many people are urgently looking for alternatives," and separately said Microsoft's tuned models have outperformed OpenAI's GPT-5.5 on specific workloads at a fraction of the cost, citing work done refining models for McKinsey.

This isn't happening in isolation. TechCrunch frames it as part of a wider industry pattern: a period of aggressive AI token usage giving way to active cost-cutting, with Amazon, Uber, Meta and Accenture all named as companies making similar moves recently. Some Silicon Valley companies are reportedly even exploring Chinese AI models as cheaper alternatives, despite security concerns around knowledge distillation.

Is this a build-vs-buy story?

Largely, yes — Microsoft had been buying frontier model access from OpenAI and Anthropic, and is now shifting weight toward models it builds and tunes itself wherever the economics work out, while keeping third-party access where it still makes sense. It's the same build-vs-buy tradeoff a lot of companies face with AI tooling, just playing out at Microsoft's scale: build in-house where volume and cost justify the investment, keep buying where a third-party model still wins on quality or speed to ship.

What should teams building on Microsoft's AI products watch for?

Anyone building on Microsoft 365 Copilot, GitHub Copilot or Azure AI Foundry should expect the underlying model mix behind familiar features to keep shifting toward Microsoft's own MAI models, even where the product interface doesn't change. That can affect output quality, tone and latency in ways worth testing for, not assuming stay constant. It's also a reminder that "which model actually answers this prompt" is becoming a moving target across most major AI products, not just Microsoft's — worth checking rather than assuming when cost or output quality suddenly shifts.