Microsoft is starting to move more Copilot work onto its own AI models, a shift that could reshape how Excel, Outlook, GitHub Copilot and Teams use artificial intelligence behind the scenes.
The company’s in-house MAI models are already handling tens of thousands of requests per week in Excel and Outlook, Bloomberg reports. Those apps previously depended more heavily on AI models from OpenAI and Anthropic.
Microsoft is bringing more Copilot work in-house
The transition is still early. According to the source, Microsoft’s MAI models process only a small fraction of total requests today. But the direction is clear: Microsoft wants to keep reducing what it spends on third-party AI over time.
That matters because Copilot is not a single product running one model in one place. The source names several Microsoft products where in-house AI is becoming part of the picture:
- Excel, where MAI models are already processing requests.
- Outlook, where MAI models are also already in use.
- GitHub Copilot, where MAI models are available.
- Teams, where a proprietary transcription model is expected to ship soon.
For users, the visible interface may not change much at first. The more important change is under the surface: which model answers a request, how much it costs Microsoft to run, and whether the result is as capable as what customers previously expected from OpenAI or Anthropic systems.
The cost logic is explicit
Microsoft’s motivation is not hard to read. The company is trying to lower its dependence on outside AI providers and reduce the cost of serving Copilot features at scale.
Mustafa Suleyman, Microsoft’s head of AI, stated the goal directly in June: "We pay a lot of money to Anthropic—so our goal is to reduce and ultimately eliminate that cost."
That statement puts the Copilot shift in plain business terms. If Microsoft can route more requests through models it owns, it can spend less on OpenAI and Anthropic. The tradeoff is whether customers receive the same level of capability when the default model changes.
The source also notes a broader tension in Microsoft’s positioning. Microsoft has argued that vendor lock-in with OpenAI and Anthropic is a bad thing and that it wants to be a platform-neutral alternative. At the same time, replacing more third-party AI with Microsoft’s own models gives the company more control over the stack customers use.
Capability remains the central question
At the Build conference, Microsoft unveiled seven new AI models, including MAI-Thinking 1, its first reasoning model. Microsoft said MAI-Thinking 1 could match Sonnet 4.6 and Opus 4.6 in coding based on human evaluations.
But the benchmarks released at the time showed a different picture. According to the source, Thinking-1 trailed OpenAI and Anthropic models by a wide margin and landed roughly on par with Deepseek V3.2.
That difference is important for Copilot customers because model choice is not only a cost issue. It can affect the quality of answers, the reliability of coding help, and the usefulness of AI features in office applications. If a cheaper in-house model becomes the default, customers may want to know whether they are receiving the same capability they associated with earlier Copilot experiences.
The source frames the risk plainly: Copilot and Office customers could end up paying the same amount while getting weaker AI, allowing Microsoft to reduce its own costs. That possibility does not mean every Copilot task will become worse. It does mean Microsoft’s model-routing choices could become more important to customers than they appear in the product interface.
Pricing could become more model-dependent
Microsoft CEO Satya Nadella has also hinted that AI billing could move further toward usage-based pricing instead of flat-rate subscriptions. The source describes one possible setup: Microsoft could make cheaper MAI models the default, while offering third-party OpenAI or Anthropic models as premium add-ons for an extra cost.
That would change how customers think about Copilot pricing. Instead of paying for a broad AI assistant and assuming the strongest available model is included, customers could face a more layered system. Basic access might rely on Microsoft’s own models, while access to outside models could come with a surcharge.
In that scenario, Microsoft would shift more of the OpenAI and Anthropic expense directly to customers who want those models. The company would still be able to advertise a Copilot product, while reserving higher-cost third-party model access for customers willing to pay more.
The training-data claim has a caveat
Microsoft has also said its MAI models are trained on clean, commercially licensed data, presenting them as safe for businesses. But the source points to a complication in the technical paper: Microsoft used the Common Crawl dataset.
Common Crawl is described in the source as a collection of freely accessible web data whose use for AI training is not legally settled. The article also notes that every other AI company does the same thing, while Microsoft portrays its own training data as particularly clean.
For business customers, that creates another point to evaluate. Microsoft’s argument is not only that in-house models can reduce costs. It is also that these models can be positioned as business-safe. The Common Crawl detail makes that claim less simple than the marketing language suggests.
The larger picture is that Copilot is entering a more complicated phase. Microsoft is not abandoning third-party AI, but it is clearly working to reduce its reliance on OpenAI and Anthropic where it can. Customers may soon need to pay closer attention to which AI model is powering which feature, what level of quality they are getting, and whether the strongest options remain included in the base price.