Microsoft MAI-Thinking-1
The flagship of seven in-house models Microsoft AI launched at Build 2026: its first reasoning model (97% AIME 25, 53% SWE-Bench Pro, 'preferred to Sonnet 4.6' in blind evals), alongside MAI-Code-1-Flash in GitHub Copilot, image, transcription, and voice models. Microsoft also introduced 'Frontier Tuning' (reinforcement learning in real-world environments) and, in early July, a $2.5B enterprise AI deployment company.
Shipped via Microsoft Foundry and inside Microsoft products, plus third-party availability on OpenRouter, Fireworks, and Baseten — with developer weight-tuning offered for the first time.
Reality review pending — we publish the executive verdict after the launch has had time to meet the real world.
Microsoft is deliberately reducing dependence on OpenAI by shipping competitive homegrown models directly into Copilot, Excel, and Foundry. For enterprises on Microsoft contracts, the default model under familiar products is quietly becoming Microsoft's own — which changes both pricing leverage and the OpenAI-Microsoft dynamic.