Microsoft is quietly weaning itself off OpenAI and Anthropic. According to TechCrunch AI, citing a Bloomberg report published Tuesday, the company has started routing a share of user prompts inside Excel and Word to its own in-house MAI models instead of the third-party software it once bragged about. This is a notable shift for a company that spent the last two years positioning itself as OpenAI’s closest partner and biggest backer.
What stands out here is the target. Excel and Word aren’t side experiments. They’re two of the most-used programs on the planet, and Microsoft is now handling some of their AI requests with models it built and owns.
What actually changed
For a while, Microsoft advertised that large parts of Office 365 ran on models from both OpenAI and Anthropic. That’s still true. The company hasn’t ripped anything out. What it’s done, per TechCrunch AI, is start answering a certain percentage of prompts with its homegrown MAI models rather than paying to send every request to an outside provider.
The groundwork was already visible. At its annual Build conference last month, Microsoft announced seven new MAI models, including an agentic coder and a text-to-image generator. When TechCrunch reached out for comment, Microsoft said it had nothing further to share.
Why this matters
Running a query through someone else’s frontier model costs money every single time. Multiply that by the hundreds of millions of people in Word and Excel, and the bill gets enormous fast. By serving even a slice of those prompts with its own models, Microsoft keeps more of that spend in-house and loosens its dependence on a partner it also competes with.
There’s strategic weight here too. Microsoft has poured billions into OpenAI, but building credible in-house models gives it leverage, a fallback, and a cheaper option all at once. Owning the model means owning the cost curve.
The bigger trend
Microsoft isn’t moving alone. TechCrunch AI frames this as part of a broader belt-tightening across the industry. Earlier this year brought a wave of what the piece calls “tokenmaxxing,” companies spending freely on AI usage. The last few months flipped the mood. Now the news cycle is full of firms getting thrifty.
According to the report, other large players have made similar moves to curb AI spending:
- Amazon
- Uber
- Meta
- Accenture
The sticker shock has gotten severe enough that some Silicon Valley companies are reportedly eyeing Chinese models for cheaper agentic solutions, even with security concerns hanging over that option. When teams start weighing potential data risk against the invoice, you know the invoice has gotten painful.
What to watch next
A few things worth tracking if you build with or buy AI tools:
- Model routing becomes standard. Expect more vendors to quietly split traffic between premium third-party models and cheaper in-house ones, sending only the hardest prompts to the expensive option.
- Quality questions. The open question is whether users notice a difference when MAI answers instead of GPT or Claude. If the swap is invisible, more of it is coming.
- Pressure on model providers. If Microsoft can offload even a fraction of Office traffic, that’s real revenue that stops flowing to OpenAI and Anthropic. Other big customers are watching.
- The China factor. Cheaper models from Chinese labs will keep tempting cost-conscious buyers, and the security debate around them is only going to get louder.
The headline story of AI in 2025 was capability. The story shaping up now is cost discipline. Microsoft using its own models inside Word and Excel is a clear signal that even the companies with the deepest pockets have decided the current spend isn’t sustainable. For everyone else, that’s a preview of the questions your own finance team is about to start asking.
More detail is available in the original TechCrunch AI report.