Microsoft Unveils Seven In-House Models to Cut OpenAI Reliance

Microsoft used its Build conference this week to show the world it no longer wants to depend on anyone else's frontier models. The company unveiled seven AI systems built entirely in-house under its MAI brand, spanning reasoning, coding, image generation, transcription, and voice. The headliner is MAI-Thinking-1, Microsoft's first reasoning model, alongside a fast coding model called MAI-Code-1-Flash.

Mustafa Suleyman, who runs Microsoft AI, framed the release around cost rather than raw capability. After tuning its models against the real workloads of consulting firm McKinsey, he said, Microsoft was able to beat OpenAI's GPT-5.5 on the same tasks with roughly ten times better cost efficiency. The models are pitched to run cheaply on Azure and land in Foundry, where developers can build on them directly.

Why build what you already resell

The strategic logic is hard to miss. Microsoft has put around 13 billion dollars into OpenAI and another 5 billion into Anthropic, and it resells both through Azure. That worked nicely until its two biggest suppliers started looking like its biggest rivals. Owning the models outright lets Microsoft keep more of the margin on every token, pass some of the savings to developers, and stop writing cheques to companies it now competes with.

Satya Nadella gave the shift its tagline, describing it as a move from consuming a frontier model to fully participating at the frontier. Suleyman has been blunter in the past about wanting long-term self-sufficiency, and seven models in one keynote is a fairly loud way to say it. Coverage from Euronews read the launch as Microsoft squaring up to OpenAI and Anthropic at once.

A hedge, not a divorce

None of this means Microsoft is dropping OpenAI tomorrow. GPT models still power much of Copilot, and the Azure resale business is real money. What the MAI family buys Microsoft is optionality: a credible in-house alternative it can lean on when frontier-model bills climb or when a supplier relationship gets awkward. It also fits the broader story from this year's Build, where Microsoft pushed agents as the new operating layer across its products. Cheaper homegrown models are what make running armies of those agents affordable.

The claims about beating GPT-5.5 on cost will get tested in the open soon enough, once developers run the MAI models on their own problems rather than McKinsey's. That is the number worth watching.