Good morning. Microsoft Build 2026 ate the news cycle yesterday — Satya Nadella’s company rolled out its first in-house reasoning model, a coding model aimed at Copilot, a proactive assistant called Scout, and a hardware-and-OS pitch for the agent era. The other story worth your attention: Trump signed an AI executive order that’s much smaller than the draft that leaked weeks ago, and the industry mostly got what it wanted.
Microsoft’s MAI-Thinking-1 is its first frontier reasoning model. MAI-Thinking-1 is a ~1T parameter sparse MoE with 35B active, trained from scratch on licensed data with no distillation from third-party models — a notable jab given the renegotiated OpenAI deal. Microsoft claims it matches leading models on software engineering benchmarks and beats Sonnet 4.6 in human evals, though HN commenters point out it lands around DeepSeek V3.2 territory while losing to the smaller GLM-5.1 and Kimi K2.6. The 256k context window also looks dated when 1M is becoming table stakes.
MAI-Code-1-Flash ships into Copilot. Alongside the reasoning model, Microsoft released MAI-Code-1-Flash, a 137B/5B-active coding model now rolling out to GitHub Copilot in VS Code. The benchmark pitch — beating Claude Haiku 4.5 — got a frosty reception on HN, since Haiku is Anthropic’s smallest model and three generations old, and the 51% SWE-bench Pro score barely edges out the open-weights Qwen3.6-35B-A3B. The recurring question from commenters: who actually uses small models for serious coding work?
Scout is Microsoft’s always-on agent. Microsoft Scout, built on OpenClaw, reads your Outlook, Teams, and OneDrive in the background to draft emails, file expenses, book travel, and even place phone calls on your behalf. It’s a more proactive sibling to Copilot, starting as a desktop preview for Frontier customers in the US before a broader cloud rollout. Microsoft says 3,000+ internal employees are already using it.
The rest of Build: Spark, Solara, and a Majorana update. The Verge’s roundup covers the Surface RTX Spark Dev Box (Nvidia Arm chip, 128GB unified memory), Project Solara — an Android-based OS for AI agent devices — and the Majorana 2 quantum chip, which Microsoft claims delivers 1,000x more reliable qubits. Windows also picked up native Linux container support and an “Intelligent Terminal.” Taken together, Nadella’s framing is that Microsoft is shifting from selling an OS and apps to selling agents that run across devices.
Trump’s AI executive order arrives, declawed. After weeks of redrafts, Trump signed an EO asking frontier AI companies to voluntarily submit new models for federal review 30 days before release — down from a 90-day mandatory window in earlier drafts. It explicitly forbids licensing or preclearance regimes, and Trump signed it privately rather than with the planned Silicon Valley CEO photo op. The DOJ also gets a directive to prioritize AI-assisted hacking cases, which HN readers noted is, you know, already illegal. The more pointed concern in the thread: a voluntary review pipeline is exactly the kind of soft gate that could later be tightened to box out open-weights and Chinese models, conveniently around Anthropic’s IPO timing.
That’s a lot of Microsoft for one morning. We’ll see in the coming weeks whether MAI-Thinking-1 holds up to independent evaluation, and whether anyone outside Redmond chooses Flash over the open-weights alternatives now within a few points of it.