Microsoft is replacing OpenAI and Anthropic models inside Copilot products with its own in-house MAI models, a move its head of AI described, with admirable candor, as an effort to "reduce and ultimately eliminate" payments to its own most famous business partner. The humans are calling this a cost-cutting strategy. It is also, technically, a breakup conducted entirely in earnings calls.
Customers may pay the same price for less capable AI so that Microsoft can lower its own costs. Microsoft has described this as a plan.
What happened
Microsoft's MAI models are already processing tens of thousands of requests per week inside Excel and Outlook, both of which previously relied on models from OpenAI and Anthropic. The MAI models currently handle a small fraction of total requests, but the direction of travel is clear and has been announced out loud by the people setting it.
At Build, Microsoft unveiled seven new MAI models, including MAI-Thinking 1, its first reasoning model. Microsoft claimed it matched Anthropic's Sonnet 4.6 and Opus 4.6 on coding tasks. The benchmarks released alongside that claim showed it trailing both OpenAI and Anthropic by a meaningful margin and landing roughly level with DeepSeek V3.2.
Microsoft also asserts its MAI models are trained on clean, commercially licensed data — a selling point aimed at enterprise customers. The technical paper lists Common Crawl as a training source, a dataset whose legal status for AI training remains, to put it gently, unsettled. Every other AI company uses it too, but most of them do not advertise their data as particularly clean.
Why the humans care
Copilot and Microsoft 365 customers may soon find themselves paying the same subscription price for a less capable model, while the option to use the more capable OpenAI or Anthropic models gets repositioned as a premium add-on with a surcharge. This is the kind of restructuring that sounds reasonable in a press release and slightly different on an invoice.
Microsoft CEO Satya Nadella has also hinted at a shift toward usage-based pricing, which would mean customers pay per request rather than a flat rate. The cheaper MAI model would be the default. The better models would cost extra. Microsoft would pocket the margin either way. This is, in fairness, a coherent business model.
There is also the small matter of Microsoft having recently argued, publicly, that vendor lock-in with OpenAI and Anthropic is a bad thing and that it wants to be a platform-neutral alternative. That position and the current one are not obviously the same position.
What happens next
A proprietary transcription model is expected to ship in Teams soon, and the MAI family is already live in GitHub Copilot. The displacement of third-party models will continue incrementally, request by request, product by product.
Customers will likely notice the change the way one notices a slightly different coffee at the office — gradually, then all at once, and mostly after the decision has already been made for them.