Microsoft Is Quietly Replacing OpenAI in Its Own Apps
On Monday, Bloomberg reported something I’ve been watching for months: Microsoft is quietly swapping out OpenAI and Anthropic models for its own in-house MAI models inside Excel and Word. Not entirely — third-party models still power parts of Office 365 Copilot — but a meaningful chunk of user prompts are now being handled by Microsoft’s own technology.
This isn’t just a Microsoft story. It’s the latest signal that the AI industry has entered a cost correction nobody talks about in the keynote speeches.
Satya Nadella’s company joins a growing list of tech giants dialing back their reliance on expensive frontier models. Amazon is doing it. Meta is doing it. Uber and Accenture have made similar moves. The industry even has a word for it now — “tokenminimizing” — which tells you how far this trend has spread.
I manage IT budgets in government, so let me tell you: when the richest companies on Earth start looking for ways to spend less on AI, the rest of us should pay attention.
What Microsoft Actually Did
According to Bloomberg’s report, Microsoft began routing certain Copilot queries in Excel and Word through its MAI models — the same ones it announced at Build last month. The shift affects what Bloomberg calls “a certain percentage” of user prompts, which is corporate speak for “we’re testing the waters before going all in.”
Microsoft launched seven new MAI models at its annual Build conference in June, including an agentic coder and a text-to-image generator. The MAI-1 model was first reported over a year ago as an internal training effort to compete with Google’s Gemini and OpenAI’s GPT series. Now those models are seeing production use inside Microsoft’s own flagship products. As I covered last week, the company also launched a $2.5 billion AI deployment business with 6,000 engineers — so this cost-cutting move is paired with serious investment in how AI gets deployed at scale.
The math is simple. Every prompt routed to a Microsoft model instead of an OpenAI or Anthropic one saves the company money at the inference layer. When you’re serving hundreds of millions of Office 365 subscribers, those fractions of a cent add up to real dollars.
This Is Bigger Than Just Microsoft
What makes this moment interesting is how many companies are arriving at the same conclusion independently.
In June, The New York Times reported that Amazon was pushing its teams to minimize token usage. The Information revealed that Meta was doing the same, reportedly spending billions on inference costs. And TechCrunch broke the story that Accenture was scrambling to stop employees from maxing out AI budgets on trivial tasks.
Meanwhile, the open source AI market is having a moment. Vercel’s AI gateway dashboard shows DeepSeek V4 Flash processing just over a third of all tokens on the platform. On OpenRouter, DeepSeek V4 Flash handles 5.3 trillion tokens weekly — more than double what the most popular frontier model manages. But here’s the kicker: when you look at spend rather than volume, Anthropic still accounts for over half of all AI spending on Vercel’s platform.
TechCrunch’s Russell Brandom captured this tension perfectly in his piece on open source AI and Anthropic. As Decagon CEO Jesse Zhang put it: “The frontier labs will keep owning discovery. Open source will increasingly own production.”
That’s a two-tier AI economy I can get behind — expensive frontier models for exploring what’s possible, cheaper open source alternatives for running what actually works.
The Developer Perspective
From where I sit as a developer and IT manager, this correction is overdue.
For the past two years, every AI discussion started with “which model is best?” and ended with “let’s just use the most powerful one.” Nobody asked whether the task actually needed a $1.37/million-token model when a 6-cent alternative would do the same job. As I explored in my piece on better models breaking workflows, even the smartest models can be the wrong tool for the job depending on context.
We’re finally asking that question. And the answers are making everyone smarter about how they deploy AI.
A few things I’ve learned watching this unfold:
- Not every task needs a frontier model. Summarizing an email chain, generating a spreadsheet formula, or rewriting a paragraph — these are properly handled by smaller, cheaper models. Microsoft seems to agree.
- The vendor lock-in risk is real. If you build your entire workflow around a single provider’s API, you’re one price hike or policy change away from a painful migration. Microsoft hedging with its own MAI models is the smart play. The Alibaba Claude Code ban showed how quickly geopolitical tensions can disrupt AI toolchains too — another reason to diversify.
- Cost visibility matters. Most teams I talk to have no idea what their AI spend actually looks like split by task type. You can’t optimize what you don’t measure.
The irony isn’t lost on me: the same companies that spent the last two years telling us AI would change everything are now quietly changing their own AI spending habits. The technology is still transformative. But the economics are getting more realistic by the quarter.
What Comes Next
If this trend continues — and I believe it will — we’re heading toward a more layered AI ecosystem. Frontier models from Anthropic, OpenAI, and Google will remain the tools for complex reasoning, research, and novel problem-solving. Cheaper, specialized models — whether from Microsoft, open source communities, or Chinese labs like DeepSeek — will handle the bulk of production workloads.
This is good for everyone. Developers get more choice at better prices. Enterprises get sustainable costs. And the frontier labs keep pushing the ceiling higher, funded by the premium they charge for the hardest problems.
Microsoft’s move is just one data point in a larger story. But it tells me the AI industry is growing up. The hype phase is giving way to the build phase. And the builders are learning to be more careful with their resources — securing their AI supply chains and thinking about long-term sustainability — which, honestly, is how real, lasting technology gets made.