Microsoft Build 2026: A Homegrown MAI Model Blitz, an “IQ” Layer for Agents, and a Quiet Goodbye to OpenAI Dependency

Microsoft opened Build 2026 in San Francisco on June 2 with its clearest declaration of independence yet. The star of Satya Nadella’s keynote was not a Copilot feature or an Azure tier — it was MAI-Thinking-1, Microsoft’s first reasoning model trained entirely from scratch, headlining a five-model barrage of homegrown AI that signals the company no longer intends to rent its intelligence from OpenAI.

The numbers Microsoft put behind MAI-Thinking-1 are aimed directly at the frontier: a mid-sized model with 35 billion active parameters and a 256K context window, tuned for multi-step instructions, long-context reasoning and code generation at deliberately low token cost. Microsoft claims independent raters preferred it to Claude Sonnet 4.6 in blind tests and that it matches Opus 4.6 on SWE-Bench Pro coding — bold claims for a model small enough to run cheaply, now in private preview on Foundry. Around it ship MAI-Code-1 (tuned for GitHub, live in Copilot and VS Code), MAI-Image-2.5 (ranked #3 for text-to-image on the Arena leaderboard, rolling into PowerPoint and OneDrive), MAI-Voice-2 and MAI-Transcribe 1.5, which covers 43 languages.

The second pillar was context. Microsoft IQ, now generally available across GitHub Copilot, Foundry and Copilot Studio, is a layer that feeds agents what they have always lacked: Work IQ pipes in organizational knowledge from Microsoft 365 signals (GA June 16), Fabric IQ grounds them in structured business data, and Web IQ gives them AI-first web search Microsoft claims runs at nearly 2.5x the speed of the next best alternative — model-agnostic and MCP-native. On top of it sits Microsoft Scout, a personal work agent that preps meetings and clears routine tasks proactively.

For developers, the GitHub Copilot desktop app entered preview — a native cockpit for orchestrating multiple agent sessions from an idea, issue or pull request — while Windows gained OS-enforced sandboxing for agents via Microsoft Execution Containers, WSL 2 picked up native GPU passthrough with CUDA support, and a Surface RTX Spark Dev Box built on Nvidia’s new superchip promises to run 120-billion-parameter models locally on 128GB of unified memory later this year. Engadget’s live coverage also caught Project Solara, an agent-first platform shown with two concept devices.

Quieter but consequential: a governance and security stack for the agent era. Agent 365 extends a single control plane over local agents through Entra, Defender and Purview; ASSERT and the Agent Control Specification arrive as open-source safety projects; and a system codenamed MDASH deploys more than 100 agents to hunt vulnerabilities — Microsoft’s answer to Anthropic’s Project Glasswing. The keynote even found room for Majorana 2, a quantum chip Microsoft says achieves a 20-second average qubit lifetime with 1,000x the reliability of its predecessor.