
Microsoft Builds Its Own Reasoning and Coding Models, Reducing Reliance on External Partners
Microsoft released MAI-Thinking-1 and MAI-Code-1-Flash on June 2, 2026, both trained entirely in-house without distillation from third-party models and using clean, commercially licensed data. The reasoning model achieves 97% accuracy on AIME 2025 benchmarks; the coding model outperforms Claude Haiku 4.5 on efficiency metrics. The shift mirrors cloud providers' historical move from reselling external software to building internal alternatives, signaling Microsoft's strategy to own its AI stack end-to-end.
Published