Frontier Models
Microsoft MAI-Thinking-1 Matches Claude Opus 4.6 on Coding Benchmarks at Mid-Size
The new reasoning model from Microsoft AI uses clean licensed data and a sparse MoE design to achieve high performance on software engineering and math tasks without third-party distillation.
MAI-Thinking-1 is Microsoft AI's first in-house advanced reasoning model released as part of a seven-model family at Build 2026. The system features a 35B active parameter sparse Mixture of Experts design with approximately 1T total parameters. It incorporates a 256k token context window. The model was announced on June 2, 2026.
What technical architecture supports MAI-Thinking-1 performance?
The architecture relies on a sparse Mixture of Experts setup that activates 35 billion parameters during inference while maintaining a total parameter count near 1 trillion. This design contributes to a smaller inference footprint compared to much larger models. The 256k token context window enables handling of extended inputs in coding and reasoning tasks.
Training occurred from scratch using clean, traceable, commercially licensed data. No distillation from third-party models occurred during development. The approach emphasizes transparency and avoids opaque data sources from other labs.
Which benchmarks demonstrate MAI-Thinking-1 capabilities?
MAI-Thinking-1 achieved a score of 52.8 percent on SWE-Bench Pro. This result positions the model as competitive with Claude Opus 4.6 on software engineering benchmarks. The performance reflects strong coding capabilities suitable for daily use.
The model recorded 97.0 percent on AIME 2025. This score serves as a primary indicator of advanced mathematical reasoning. Additional results include 87.7 percent on LiveCodeBench v6.
| Model | SWE-Bench Pro | AIME 2025 | Active Parameters |
|---|---|---|---|
| MAI-Thinking-1 | 52.8% | 97.0% | 35B |
| Claude Opus 4.6 | Competitive | Not reported | Larger |
| Claude Sonnet 4.6 | Not reported | Not reported | Not reported |
How do human evaluations position MAI-Thinking-1 against peers?
Independent human raters on Surge preferred MAI-Thinking-1 for overall quality in blind side-by-side evaluations to Claude Sonnet 4.6. The preference highlights strengths in practical output quality beyond automated benchmarks.
The smaller inference footprint allows deployment in scenarios where larger models prove impractical. This efficiency supports broader accessibility for developers and enterprises.
What availability options exist for MAI-Thinking-1?
The model entered private preview through Microsoft Foundry. Expansion to additional regions remains planned. Access will also extend to the MAI Playground in the future.
What statements did leadership make regarding the model?
Microsoft AI CEO Mustafa Suleyman addressed the model's standing relative to other systems in public comments.
It’s now roughly on par with Opus 4.6, at least on the benchmarks. We haven’t deployed it at scale into production, so there’s still lots more work to do there. But it’s an extremely strong reasoner and scored 97 percent on AIME, which is the primary measure for its reasoning performance, at least on the benchmarks.Mustafa Suleyman, Microsoft AI CEO
What implications arise for enterprise AI development?
The clean data training process may set precedents for traceable model development in the industry. Enterprises gain access to a model that delivers frontier-level coding performance without dependence on external distillation.
The design supports daily coding workflows due to its efficient footprint. Organizations can integrate the model into existing Microsoft ecosystems via Foundry.
What steps follow the initial release of MAI-Thinking-1?
Further scaling and production deployment testing form the immediate priorities. The model belongs to a seven-model MAI family that continues to expand.
- Scale deployment testing in production environments.
- Expand regional availability beyond current preview.
- Integrate additional features into the MAI Playground.
- Release further models from the seven-model MAI family.
Frequently asked
What is the active parameter count of MAI-Thinking-1?
The model has 35 billion active parameters as part of its sparse Mixture of Experts architecture with roughly 1 trillion total parameters.
On which benchmark did MAI-Thinking-1 score 52.8 percent?
It scored 52.8 percent on SWE-Bench Pro per the Microsoft AI report.
Is MAI-Thinking-1 available for public use?
It is currently in private preview through Microsoft Foundry with plans for broader access.