You’re reading an issue of “The AI Economy,” my newsletter exploring the forces shaping the AI era—tracking how AI is rewriting business, work, technology, and culture. Subscribe to get expert insights and curated updates delivered straight to your inbox.
For most of the generative AI era, Microsoft has been the most prominent buyer of someone else’s intelligence. Its Copilot stack runs on OpenAI’s GPT models. Azure customers reach for frontier reasoning through the partnership. The arrangement has worked, and it has also been a structural dependency that the company has spent the last year visibly working to reduce.
At its Build developer conference this week, Microsoft took another step towards that independence, launching seven new in-house models from its AI Superintelligence Team. The headliner is MAI-Thinking-1, Microsoft’s first reasoning model, that the company says matches Anthropic’s Opus 4.6 model in coding according to its own testing, and is designed for complex multi-step instructions, long-context reasoning, and code generation, all at what the company calls a low token cost.
In addition to MAI-Thinking-1, Microsoft is releasing new versions of its homegrown image generation, transcription, and voice models. It’s also debuting a new coding model that could rival coding models from OpenAI and Anthropic.
These seven models are part of the company’s push to create what it calls “humanist superintelligence,” state-of-the-art AI with capabilities “explicitly designed” to serve people and organizations rather than replace them. “The type of AI that we create really does matter,” Mustafa Suleyman, the chief executive of Microsoft AI, proclaimed on the Build stage. “We need an AI that places humanity first, that always prioritizes human well-being and human progress.”
He added that the MAI models “are all built with real attention to detail and a commitment to making very practical and efficient tools that are tuned to just how you work in the real world.”
Microsoft’s First Homegrown Reasoning Model
MAI-Thinking-1 is a mid-sized reasoning model with 35 billion active parameters and a 128K context window. Microsoft says it was trained from scratch, with zero distillation, on enterprise-grade, clean, and commercially licensed data. In other words, the model’s capabilities are its own rather than borrowed from another lab’s outputs, such as OpenAI, and the training corpus is the kind Microsoft can defend if a copyright claim arrives.
Through reasoning, AI models can work through complex problems step by step rather than making a prediction in a single pass. With MAI-Thinking-1, Microsoft joins a field that has gotten crowded fast. OpenAI, Anthropic, Google, xAI, DeepSeek and Alibaba have all shipped reasoning models since OpenAI introduced the form factor with o1 in late 2024. Microsoft was the conspicuous absence in that group; today it isn’t.
Suleyman noted that MAI-Thinking-1 was preferred by independent human raters on search over Anthropic’s Sonnet 4.6. “It’s achieved 97 percent on AIME 2025, which is obviously the key measure of its general-purpose reasoning abilities, but most importantly of all, it’s now at 53 percent on the SWE-bench Pro, which places it right alongside Opus 4.6, at least on the toughest coding benchmark that’s out there.”
MAI-Thinking-1 is now available in Microsoft Foundry in private preview.
The Rest of the MAI Model Slate
Beyond MAI-Thinking-1, Microsoft used Build to update four model families it has been building out since last summer.
To start, Microsoft is broadening MAI-Image-2.5’s reach. The base model launched a week ago, ranked third on Arena’s text-to-image leaderboard, with Microsoft citing gains in text rendering and commercial imagery. Today’s news adds a flash variant tuned for speed and cost, the family’s first move into image-to-image work, and availability across PowerPoint, OneDrive and Foundry. “Flash is here for super-efficient production workloads, while 2.5 gives you that maximum fidelity and professional-grade performance,” Suleyman said. Microsoft also says the model surpasses Google’s Nano Banana Pro on ELO.
The pace behind that announcement is worth noting. Microsoft introduced MAI-Image-1 earlier this year, debuting in the top 10 of LMArena’s text-to-image leaderboard and shipping into Bing Image Creator and Copilot. MAI-Image-2 followed in March, with an emphasis on photorealism and readable in-image text. In April, the company released MAI-Image-2-Efficient, a faster, cheaper variant for high-volume production workloads.
Next, Microsoft is extending MAI-Transcribe’s footprint. The base model launched on Foundry on April 2, supporting the 25 most-used languages at $0.36 per hour of audio and delivering 2.5 times faster performance than Microsoft’s Azure Fast transcription for batch workloads. Today’s update brings coverage to 43 languages, with streaming support promised in a later release. Suleyman boasted that MAI-Transcribe-1.5 outperforms Google’s Gemini and OpenAI’s flagship transcription models. Microsoft has integrated the new model inside GitHub, Teams, Copilot, and the Dynamics 365 contact center. It’s also available inside Foundry.
MAI-Voice is perhaps Microsoft’s oldest model. It debuted in August 2025 in Copilot Daily, Podcasts, and Copilot Labs, capable of generating a full minute of audio in under a second on a single GPU, before reaching commercial availability on Foundry in April. “It has beautiful prosody, natural-sounding delivery, fine-grain emotional control, and it’s available in 15 languages, with many more coming soon,” Suleyman said. Microsoft has also unveiled Voice to Flash, which supports ultra-latency-sensitive voice agents, something that’s “the big thing in 2026.”
Meet Microsoft’s Coding Model
Microsoft owns one of the most widely used developer platforms on the market, GitHub, along with GitHub Copilot, the most widely deployed AI coding assistant. Since Copilot launched in 2021, it has run primarily on OpenAI’s models, with Claude and other options added to its model picker over the past year.
Today, Microsoft is putting a first-party option into that stack with MAI-Code-1, a five-billion-parameter coding model tuned for GitHub and built to run quickly and cheaply. Microsoft says it’s now available in Copilot and VS Code, though it didn’t specify whether the Copilot reference here means GitHub Copilot, the developer assistant, or the so-called super Copilot app. Whichever it turns out to be, MAI-Code-1 gives Microsoft something it has lacked for the entire history of its AI coding work: a model it owns running inside the products that defined AI-assisted coding.
The timing of an ‘ultra-efficient’ coding model is hard to read as anything other than intentional. On the day before Build, GitHub moved Copilot to usage-based billing, replacing flat-rate subscriptions with AI credits consumed by token use. Developer reaction has been sharp; some users have shared projections of 10x to 50x increases on agentic workflows, and TechCrunch declared the end of Copilot’s ‘golden age.’ A model designed to consume fewer tokens per request is the most direct technical answer Microsoft can offer to a problem its own pricing change just made acute. Whether MAI-Code-1 actually delivers on that depends on numbers Microsoft hasn’t published. The positioning, though, is unambiguous.
It’s worth asking, given the visible reshaping of the Microsoft-OpenAI relationship, why today’s slate doesn’t include a head-on competitor to GPT-5, Claude or Gemini at the general-purpose flagship tier. The likely answer is enterprise demand. These organizations already have access to a flagship LLM, and dislodging an entrenched default inside the same buyer can be an expensive, low-margin fight. Instead, Microsoft’s strategy is about differentiation by category: a reasoning model, a coding model, refreshes across image, voice and transcription. Build the models for categories where the market hasn’t settled on a default. Then, over time, as organizations embrace Microsoft’s humanist AI philosophy, they may want to use its general-purpose LLM whenever that is released.
Disclosure: I attended Microsoft Build as a guest of the company, with my travel and expenses paid for. However, what I write reflects my own reporting and analysis. No one reviewed or approved this piece before publication.
