MAI-Thinking-1
Microsoft's first in-house reasoning model, a sparse MoE with ~1T total and 35B active parameters and 256K context, trained without third-party distillation.
Microsoft says it matches Claude Opus 4.6 on SWE-Bench Pro and scores 97.0% on AIME 2025, preferred to Claude Sonnet 4.6 in blind human evals. Announced at Build with the other MAI models; public preview in Foundry followed 2026-08-12.
- Date
- Tuesday, 2 June 2026
- Lab
- Microsoft
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| AIME 2025 / AIME 2026 | 97.0% / 94.5% also matches Claude Opus 4.6 on SWE-Bench Pro | company |
| Blind preference vs Claude Sonnet 4.6 | preferred 1,276 tasks, single and multi-turn | company |
All benchmark claims are Microsoft's own; no independent evaluation opened. Thurrott describes the Build launch as private preview; the dedicated post is dated 2026-08-12.
Sources
- microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/
- microsoft.ai/news/introducing-mai-thinking-1/
- www.thurrott.com/a-i/336960/build-2026-microsoft-launches-first-flagship-reasoning-ai-mode
This record was checked against its sources on 6 October 2026. How we check