Item: Microsoft
Author: Karl Strauchman

Microsoft opened Build 2026 at Fort Mason in San Francisco on June 2 by shipping seven in-house MAI models, headlined by MAI-Thinking-1, a sparse Mixture-of-Experts reasoner with 35 billion active parameters and roughly a trillion total, now in private preview on Microsoft Foundry. The Microsoft AI Superintelligence Team is calling it "mid-sized" and built for low token cost. It's also, by Microsoft's own accounting, the first frontier-class reasoning model the company has trained without distilling from a partner's outputs.

The framing onstage was unsubtle. "We believe the time has come for every company to just move from consuming a frontier model to fully participating at the frontier in the frontier ecosystem," Satya Nadella said, which is a polite way of describing a company that has put $13 billion into OpenAI and another $5 billion into Anthropic announcing that it would now prefer to be a peer.

The specs do real work. MAI-Thinking-1 ships with a 256,000-token context window, enough to ingest a 600-page document in a single pass. Microsoft says blind raters at Surge, its independent rating partner, preferred its outputs to Claude Sonnet 4.6, and that on SWE-Bench Pro it matches Claude Opus 4.6, the model Anthropic released in February. Those numbers sit in footnotes on the official post, which is the appropriate amount of hedge for a private-preview claim, but the comparison set itself is the message.

The rest of the lineup fills out the stack. MAI-Code-1 is already live in GitHub Copilot and VS Code, with a MAI-Code-1-Flash variant alongside it. MAI-Image-2.5 lands in PowerPoint today, claiming third on Arena AI's text-to-image leaderboard and second on image-to-image, ahead of Google's Nano Banana 2. MAI-Transcribe-1.5 covers 43 languages; MAI-Voice-2 and its Flash variant add more than 15 more.

Foundry itself is the strategic object. Microsoft is distributing the catalog through Fireworks AI, Baseten, and OpenRouter, which is how you behave when you want to be a model provider rather than a model reseller. The timing isn't incidental: Anthropic confidentially filed for an IPO on June 1, and CNBC reports both Anthropic and OpenAI are moving toward public offerings. Microsoft's two biggest model investments are about to become entities with their own shareholders, their own pricing power, and their own incentives to disintermediate Azure.

Shipping its own reasoning tier the day after one of those filings isn't a coincidence. It's the structural answer to a dependency Microsoft has spent three years insisting wasn't one.

Microsoft ships seven in-house MAI models at Build, led by a 35B-active MoE reasoner trained without distillation

Sources