Microsoft just announced seven AI models it built itself. Not borrowed, not licensed from OpenAI or anyone else. Built from scratch by its own AI Superintelligence team. And if you’ve used PowerPoint’s image generation feature recently, there’s a good chance you’ve already encountered one of them without realising it.

The models all carry the MAI name. MAI stands for Microsoft AI, and it is the brand Microsoft is now using for models it develops internally rather than sourcing externally. For years, Microsoft’s identity in artificial intelligence was tied primarily to its large-scale investment in OpenAI, but Build 2026 marks the clearest signal yet that the company wants to own the models powering its own products.

At Build 2026, Microsoft announced a family of seven in-house models under the MAI brand, developed by the company’s AI Superintelligence team, covering the full range of use cases from software development to voice and image generation.

Here is what each one does.

The Seven MAI Models, Explained Simply

The headline model is MAI-Thinking-1, a 35 billion parameter reasoning model with a 256K context window, built without distillation from other models. A reasoning model is one that works through problems step by step before giving you an answer, similar to how OpenAI’s o-series models work or how Claude handles complex tasks. In blind evaluations, MAI-Thinking-1 was favoured over Claude Sonnet 4.6 and performed on par with Claude Opus 4.6 on the SWE Bench Pro coding benchmark, according to Microsoft. These are Microsoft’s own claims, so treat them with appropriate scepticism, but the numbers are competitive enough to take seriously.

MAI-Code-1-Flash is tuned for fast, low-cost coding assistance and is already rolling out in VS Code through the GitHub Copilot model picker. If you pay for GitHub Copilot, this is the most immediate way a regular person might encounter a MAI model directly.

MAI-Image-2.5, and its faster Flash variant, is a text-to-image model now available inside PowerPoint and coming soon to OneDrive. It currently ranks third on arena.ai’s image model leaderboard, a ranking driven entirely by human preference votes rather than internal benchmarks. That is a meaningful result worth noting.

The remaining models cover voice and transcription. MAI-Voice-2 handles speech generation across multiple languages, and MAI-Transcribe-1.5 delivers transcription across 25 languages, with batch transcription speed 2.5 times faster than Microsoft’s existing Azure Fast offering.

So Can Regular People Actually Use These?

This is the honest part. These models are not set up like ChatGPT or Claude where you create an account, pick a plan, and start chatting. The access story is more layered than that.

Developers can access MAI models through Microsoft Foundry and try them in the MAI Playground, though the Playground is currently US only. Those without Foundry access can fill out a request form on the Microsoft AI website.

For individual creators or smaller teams who want quick, no-setup access, there is real friction right now. The models are not broadly available through consumer-facing tools the way Midjourney or DALL-E are.

The exception is what is already embedded in Microsoft products you may already use. MAI-Image-2.5 is live in PowerPoint. MAI-Code-1-Flash is arriving in VS Code. Microsoft also confirmed that MAI models will be accessible via third-party platforms including Fireworks AI, Baseten, and OpenRouter, which will eventually make them reachable without an Azure account. That access is not fully live yet.

For most people right now, the experience will be passive. Microsoft’s own products will quietly get smarter as these models roll out behind the scenes.

What to Watch From Here

The MAI model family is a first step, not a finished product. The ones worth keeping an eye on are MAI-Image-2.5, because its arena.ai ranking suggests it is genuinely competitive with the best image models available right now, and MAI-Thinking-1, because reasoning models are where the most interesting AI development is happening across the board.

When these models land on OpenRouter and become accessible without an Azure account, that will be the moment to actually put them through their paces properly.

Frequently Asked Questions

What are the Microsoft MAI models?

MAI stands for Microsoft AI. The MAI models are a family of seven artificial intelligence models built in-house by Microsoft’s AI Superintelligence team, announced at Build 2026 in June 2026. They cover reasoning, coding, image generation, voice, and transcription tasks. They are designed to power Microsoft’s own products and be made available to developers through the Microsoft Foundry platform.

Can regular people use Microsoft’s MAI models?

Not directly in the way you would use ChatGPT or Claude right now. Most MAI models are available to developers through Microsoft Foundry or the MAI Playground, which is currently limited to US users. However, MAI-Image-2.5 is already live inside PowerPoint, and MAI-Code-1-Flash is rolling out in VS Code via GitHub Copilot. Microsoft has also confirmed that MAI models will eventually be accessible through third-party platforms like OpenRouter, which will widen access significantly.

How does MAI-Thinking-1 compare to ChatGPT or Claude?

MAI-Thinking-1 is a reasoning model, similar in design to OpenAI’s o-series or Anthropic’s Claude when handling complex tasks. Microsoft claims it matches Claude Opus 4.6 on a standard coding benchmark and that blind evaluators preferred it over Claude Sonnet 4.6. These are Microsoft’s own claims and have not been independently verified at scale yet. The model is currently in private preview and not available for public testing.

The Bigger Picture

Microsoft has spent the last few years being the company that brings you AI from somewhere else. Build 2026 is the first real signal that they want to be the company that makes it. Whether the MAI models can compete at scale with what OpenAI, Anthropic, and Google are shipping is a question that will take months to answer properly.

What is clear is that the direction has changed. And for anyone curious about where AI is heading, that shift is worth paying attention to.

If you want to follow along as these models roll out and become more accessible, sign up for the August Wheel newsletter at newsletter.augustwheel.com. I’ll flag when regular people can actually get hands on with them.