At Build 2026 on June 2, Microsoft’s most concrete AI news was not another agent platform, but seven models trained from scratch by its own superintelligence team. The lineup spans reasoning, code, image generation, transcription and voice, giving Microsoft a more complete first-party model stack.
The headline model is MAI-Code-1-Flash, a lightweight coding model designed for agent work inside VS Code and GitHub Copilot. Microsoft compared it with Claude Haiku 4.5 and said it reached 51.2% on SWE-Bench Pro versus 35.2% for Haiku, a 16-point gap, while using up to 60% fewer tokens for similar work. It also led by 28.9 points on IF Bench and scored 85.8% on an adversarial reasoning benchmark.
Microsoft stressed that MAI-Code-1-Flash was built end to end in-house using clean, licensed and compliant data. That wording matters in a market full of copyright fights over training corpora. The other models include MAI-Thinking-1 with 35 billion active parameters and a 12.8 million token context, a Flash version, MAI-Image-2.5, MAI-Transcribe-1.5 supporting 43 languages, and MAI-Voice-2 with more than 15 languages.
The strategic message is clear: Microsoft wants more than distribution rights to other companies’ models. If its own small, fast model can sit inside Copilot, VS Code, PowerPoint and AI Foundry, the company has a lower-cost lever it controls directly. Whether the benchmark gap holds in real engineering work will depend on developer use over the next months.
Sources: CocoLoop; Microsoft AI; SiliconANGLE; CNBC