Microsoft has started routing work inside its own products away from OpenAI and Anthropic, and towards models it built itself. Satya Nadella laid out the reasoning in a long post on X, describing a system where Microsoft’s in-house MAI models take over tasks once handled by frontier models from its partners. GitHub Copilot, Excel and Outlook are already showing what he called promising early results. Copilot Chat and PowerPoint are next in line.The logic is cost. Nadella’s framing is that software now has a real marginal cost for the first time, which makes the choice of model a business decision rather than a technical default. Frontier models stay in the mix for frontier problems. Everything else gets pushed to something cheaper that performs just as well. Microsoft says its MAI models already beat general-purpose frontier models on many use cases while burning a fraction of the tokens.
Why Microsoft built its own AI models instead of leaning on OpenAI and Anthropic
The MAI family arrived on June 2, when Mustafa Suleyman announced seven new models spanning reasoning, coding, image, voice and transcription. The flagship reasoning model, MAI-Thinking-1, was trained from scratch on what Microsoft calls clean, traceable data—no distillation from other labs. MAI-Code-1-Flash runs 5 billion active parameters and is built specifically for GitHub Copilot and VS Code. Microsoft claims it is comparable to Haiku but cheaper.That cheapness is the entire point. Microsoft’s tuned model for Excel matches GPT 5.4 while running up to 10 times more efficiently, according to the company. At Microsoft’s scale, across hundreds of millions of Office seats, that gap compounds fast.
What reinforcement learning environments mean for Microsoft’s enterprise customers
The mechanism Nadella keeps returning to is the reinforcement learning environment, or RLE. Rather than train a model in isolation and then bolt it onto a product, Microsoft trains models inside the actual product harness, rewarding them for finishing the tasks customers care about. Suleyman calls these training gyms. Microsoft is selling the same capability to enterprises through Foundry under the name Frontier Tuning, letting companies build models on their own workflow data.Nadella was careful about one detail. Models from OpenAI and Anthropic remain part of the orchestration system inside Microsoft products. He also set a test for whether Microsoft is genuinely in control: evaluation scores should keep improving even if any single model is pulled out. Memory, context, tools and skills sit outside the model deliberately, so no one vendor holds the leverage.For Microsoft, which has poured billions into OpenAI, that is a notable position to stake out publicly. The company is describing a template it expects other software firms to copy—use the expensive model where it earns its keep, and route the rest somewhere cheaper you own.
