Microsoft released seven AI models at its Build conference, the most it has ever shipped at once, and aimed them at the enterprise. The flagship, MAI-Thinking-1, is its first reasoning model, and Microsoft benchmarked it against Anthropic’s Claude Opus 4.6 and Sonnet 4.6. Opus 4.7 has been out since April. Opus 4.8 shipped five days before Build. So Microsoft compared its newest model to the Claude from two versions ago, which is a familiar way to announce that you have caught up, and a strange way to spend a keynote, unless catching up was never the point.

It was not the point. MAI-Thinking-1 is, by the numbers Microsoft has shared, very good: it reportedly ties Opus 4.6 on a hard coding benchmark and beat Sonnet 4.6 in a blind human test. Very good is not best, and Microsoft seems fine with that, because the model is cheap and runs on its own Azure, so it no longer has to pay Anthropic or OpenAI to serve it. The company says that after tuning its models for McKinsey, it matched OpenAI’s GPT-5.5 at a tenth of the cost. Microsoft is selling a very good model, for much less money, inside software you already pay for, and it would like you to look at the price more than the benchmark.

This is a bet about how enterprises behave, which is a subject Microsoft knows cold. Enterprises do not buy models. They buy the thing the model lives inside: the Microsoft 365 seat, the Azure contract, the Copilot license, the compliance review that already passed. A model that is ninety percent of Claude but native to all of it, on an invoice you already pay, is a much easier purchase than a better model that shows up with an integration project and a new vendor to onboard. This is where good-but-not-frontier models tend to win. Switching is work. Microsoft is betting you would rather not.

The models were trained from scratch on clean, licensed data with no distillation from anyone else’s, which Microsoft calls trust and which also, conveniently, leaves no one else to pay and no one else to blame. “Models you can trust,” wrote Mustafa Suleyman, who runs Microsoft AI. In the enterprise, trust mostly means the model will still be here next year, at last year’s price, indemnified exactly as promised. It is a boring pitch, and a good one.

Microsoft is doing this now because renting got uncomfortable. It built its whole AI position on OpenAI, and in April the two of them rewrote the deal: no more revenue share to Microsoft, OpenAI free to sell on AWS and Google Cloud, and the exclusivity that made Azure the only place to get the best models gone. Anthropic passed OpenAI on revenue by putting Claude on all three clouds, and Claude’s coding tool quietly beat Microsoft’s own Copilot. The supplier became a competitor, the competitor became a supplier, and Microsoft was reselling both of them. Building its own models is how it stops.

So Microsoft is doing what everyone who rents a critical layer eventually does. ByteDance decided it would rather build its own chips than keep renting them. CATL decided that whoever supplies the power should own the power. Microsoft has decided that the company distributing everyone else’s models should keep a few of its own, even the un-best ones, because the best model is the one you do not have to pay a competitor to use.

None of this needs MAI-Thinking-1 to beat Claude, which is fortunate, because it does not. It needs the enterprise to look at a very good model that costs less and is already installed and conclude that this is fine, that the last few IQ points of the frontier are not worth a migration, that staying beats leaving. So Microsoft built seven models, priced them low, dropped them into the apps its customers open every morning, and wrapped the whole thing around one quiet request, which is please do not leave.

–
By the Control Plane Editorial Team