Microsoft's MAI Thinking 1 introduces a 1T parameter Sparse MoE architecture, outperforming Claude Sonnet 4.6 on AIME 2025 benchmarks.

The landscape of artificial intelligence reasoning has shifted. On August 12, 2026, Microsoft officially released MAI Thinking 1, the flagship reasoning model from its dedicated Microsoft AI Superintelligence team. This release marks a pivotal moment in Microsoft's strategic evolution, signaling a move toward high-performance, first-party models designed to reduce reliance on third-party providers.
Unlike many contemporary models that rely on distillation from larger predecessors, MAI Thinking 1 was built from the ground up. It represents a pure pursuit of reasoning capability, utilizing clean, traceable, enterprise-grade data and a proprietary reinforcement learning framework. For developers and AI engineers, this means a model with more predictable behaviors and a more robust foundation for complex, multi-step logic tasks.
At the core of MAI Thinking 1 is a massive Sparse Mixture-of-Experts (MoE) architecture. While the total parameter count sits at approximately 1 trillion, the model is highly efficient during inference, activating only 35 billion parameters per token. This balance allows for deep reasoning capabilities while maintaining manageable latency for complex queries.
The model supports a massive 256K token context window, roughly equivalent to a 600-page document. This makes it an ideal candidate for long-context RAG (Retrieval-Augmented Generation) and deep analysis of large codebases. Furthermore, it is fully compatible with the Chat Completions API and features native support for function calling and granular developer instructions.
MAI Thinking 1 has demonstrated industry-leading performance across mathematical and logical reasoning benchmarks. In the AIME 2025 competition, the model achieved a score of 97.0, surpassing Claude Sonnet 4.6's 95.6. It maintained high standards in AIME 2026 with a score of 94.5, and recorded an 84.2 on the GPQA Diamond benchmark, showcasing its ability to handle PhD-level scientific reasoning.
In the realm of software engineering, the model is a formidable competitor. It scored 73.5 on SWE-bench Verified, placing it firmly in the territory of Claude Opus 4.6 (80.8). While it trails slightly behind the absolute frontier on Terminal-Bench 2.0 (46.0 vs GPT-5.1's 75.1), its performance on SWE-bench Pro (52.8) is nearly neck-and-neck with Sonnet 4.6 (53.4).
The model's intelligence was refined through a sophisticated 'hill-climbing' RL methodology. During training, the AIME 2025 pass@1 rate climbed from a mere 0.145 at step 5 to a staggering 0.967 at step 5,074 across 235 distinct checkpoints. This iterative refinement ensures that the model's reasoning pathways are optimized through direct experience rather than imitation.
Microsoft has integrated safety directly into the reasoning loop. Rather than applying safety as a post-processing layer, MAI Thinking 1 utilizes a unified RL reward structure where safety and capability are trained simultaneously. Potential harms are weighted by severity, ensuring the model learns to navigate complex ethical boundaries while maintaining high utility.
For agentic workflows, the model's coding capabilities are not just theoretical. Agentic coding performance is graded using real test suites within executable environments. This ensures that when the model generates code or executes functions, the output is verified by functional correctness rather than just linguistic plausibility.
MAI Thinking 1 is positioned as a premium reasoning model, reflecting its massive scale and the computational intensity of its MoE architecture. Pricing is structured to support both high-volume enterprise agents and specialized research tasks.
Currently, there is no free tier for the flagship reasoning model, though developers can access it via public preview on Microsoft Foundry and Baseten.
The primary strength of MAI Thinking 1 lies in complex reasoning and autonomous agentic behavior. It is exceptionally well-suited for high-stakes software engineering tasks, where it can navigate large repositories and solve intricate bugs via SWE-bench-style workflows.
Additionally, the 256K context window makes it a powerhouse for deep RAG applications, such as legal document analysis, scientific research synthesis, and comprehensive technical documentation auditing. Its ability to follow complex developer instructions makes it a top-tier choice for building autonomous coding agents.
Developers can access MAI Thinking 1 through the Microsoft Foundry public preview. For those integrated into the Microsoft ecosystem, it is also being rolled out via Azure AI services. For alternative deployment paths, the model is available on Baseten.
To begin integration, developers should use the standard Chat Completions API, ensuring that system prompts leverage the model's advanced capability for following detailed developer instructions and function-calling schemas.
API Pricing β Context: 256K