Microsoft used its Build 2026 keynote on June 2 to reset its frontier AI strategy, launching MAI-Code-1-Flash inside GitHub Copilot and previewing MAI-Thinking-1, its first reasoning model trained entirely on commercially licensed data without distillation from any third-party model. The pair anchor a seven-model in-house lineup that signals a deliberate decoupling from OpenAI in Microsoft's most-used surfaces.
MAI-Code-1-Flash Targets Copilot Workloads
MAI-Code-1-Flash is rolling out to GitHub Copilot Free, Pro, Pro+ and Max users in Visual Studio Code, starting in the model picker and the default auto picker. Microsoft trained the model directly with the same Copilot harness developers use in production, then evaluated checkpoints across software engineering, repository question answering, refactoring and telemetry-grounded tasks. The result, the company says, is a coding assistant that outperforms Claude Haiku 4.5 across SWE-Bench Verified, SWE-Bench Pro, SWE-Bench Multilingual and Terminal Bench 2 — including a 16-point lead on SWE-Bench Pro (51.2 percent vs. 35.2 percent) while using up to 60 percent fewer tokens.
MAI-Thinking-1: 35B Active Parameters, No OpenAI Data
MAI-Thinking-1 is a mid-sized sparse mixture-of-experts model with 35 billion active parameters and a 256,000-token context window, available now in private preview on Microsoft Foundry. Microsoft positions it for high-complexity reasoning, with results on SWE-Bench Pro and adversarial reasoning categories that the company says justify the new architecture. The model is part of what Microsoft AI's superintelligence team described as a hill-climbing approach rather than a moonshot — a series of focused models trained on owned data and rolled into product surfaces immediately.
Seven Models, Two Strategies, One Independence Play
The broader Build 2026 lineup includes MAI-Image-2.5 and MAI-Image-2.5 Flash for image generation and editing, MAI-Transcribe-1.5 covering 43 languages and MAI-Voice-2 for voice cloning in 15-plus languages. Stacked against Microsoft's earlier Edge Aion-1.0 on-device launch and Anthropic's frontier roadmap, the rollout positions Microsoft to substitute first-party models inside Copilot and Foundry without breaking compatibility for developers. NVIDIA's GTC Taipei announcements with TSMC on AI fabs and GTC Taipei further underline that the AI-stack is being rebuilt model-by-model and fab-by-fab — and Microsoft just made clear it intends to own the model layer for its own products, much as Nscale is owning the European sovereign AI infrastructure stack.
Reporting based on coverage from Microsoft AI, CNBC and TechTimes.