Microsoft MAI: A Strategic Break with OpenAI and Anthropic

During its annual Build 2026 conference, Microsoft made a major announcement by introducing a new family of artificial intelligence models called MAI (Microsoft AI). This initiative marks a strategic turning point for the company, which aims to reduce its reliance on OpenAI and Anthropic. The flagship model, MAI-Thinking-1, stands out with its MoE architecture and 35 billion active parameters, promising top-tier performance at a lower cost than competing models from Anthropic.

Microsoft has begun to shed its historical dependencies on OpenAI and Anthropic. At the Build conference, held on June 2, the Redmond firm presented a family of cutting-edge generative AI models. This new iteration is the result of the work of the Microsoft AI Superintelligence Team, a dedicated artificial intelligence group launched in November 2024 and led by Mustafa Suleyman, co-founder of DeepMind. Seven specialized models were announced, with Microsoft showcasing the five main ones.

A Cutting-Edge Model…

Named MAI (for Microsoft AI), the family was designed from the ground up by the company. Microsoft focused its efforts on two differentiating strengths: legal risk on one hand, and compactness on the other. The entire suite was fully trained on commercially viable quality data. The main model, MAI-Thinking-1, the company's first reasoning AI, comes with a MoE architecture and only 35 billion active parameters at inference.

Microsoft is not currently disclosing benchmark details, but the model is said to be on par with Claude Opus 4.6 on SWE-bench Pro, a benchmark test for code in agentic mode. The company has not yet communicated its pricing structure, while assuring that the model would offer top performance at a price significantly lower than Anthropic's models. The only downside is a context of 128,000 tokens, at a time when the latest top models manage at least a million. The model is already available in Microsoft Foundry.

… and Specialized Models

In parallel, the Microsoft Superintelligence team unveiled MAI-Image-2.5 and MAI-Image-2.5 Flash, an update to its leading text-to-image model positioned as a direct competitor to Gemini Nano Banana Pro. In benchmarks, it is said to outperform Google's model. Designed for creative workflows, it generates images from a prompt just as it edits them. It is now integrated into PowerPoint, OneDrive, and Foundry. Microsoft claims it offers the best quality-to-price ratio on the market.

The company also launched MAI-Transcribe-1.5, an update to its transcription model that boasts SOTA scores in over 43 languages. The team took the opportunity to launch MAI-Voice-2 and MAI-Voice-2 Flash, which generate voices from a prompt in most current languages. Finally, Microsoft released MAI-Code-1, a specialized model for rapid code editing that is highly efficient. It is available in Copilot and VS Code. The MAI models will also be natively accessible on Fireworks AI, Baseten, and OpenRouter.

Other AI Announcements from Build

Additionally, Microsoft announced new agent products. Starting with Microsoft Scout, a personal autonomous work agent (positioned similarly to Claude Cowork) available today to a panel of testers. Developed on an OpenClaw foundation and merged with Microsoft's AI agent WorkIQ, it integrates with Teams and Outlook to proactively manage all tasks in the digital workplace. Microsoft also unveiled its agent platform, Microsoft Agent Platform, which covers the entire agent lifecycle: from development in GitHub to deployment in Foundry, and exposure in M365 and Teams.

The company added Frontier Tuning, which applies reinforcement learning to adapt an agent's underlying model to the company's data and workflows, within its compliance scope (a somewhat vague concept, we agree). For these agents to be relevant, they must understand the company's context. This is the purpose of Microsoft IQ, a software layer that connects company data to agents and interfaces natively with WorkIQ, Fabric IQ, and Foundry IQ. Finally, Microsoft launched Web IQ, a web search engine designed for agents, compatible with MCP and up to 2.5 times faster than its competitors.

With this first MAI family, Microsoft is crossing a threshold. The Redmond firm is advancing into Google's territory: that of specialized and efficient models. It does not claim to compete head-on with the most powerful models on the market. Its bet lies elsewhere: more compact models tailored for specific uses, at a price that should make a difference. The benefits are twofold. Microsoft reduces its historical dependence on OpenAI and Anthropic, while also fulfilling the mission for which the Superintelligence Team was unofficially created. The goal is not to acquire new clients, but to better serve its existing ones, keeping them on its own models rather than sending them to the competition.