I Tested All 4 New Microsoft AI Models: The Brutal Truth

Microsoft pulled back the curtain on a bold new vision for Windows at Build 2026, one that looks almost nothing like the operating system of today. Near the end of its keynote, the company also introduced a fresh lineup of Microsoft AI models for image generation, reasoning, transcription, and voice. You can try them all for free right now. I did. And honestly, I’m not impressed. Here’s the real story, along with everything you should know about the new MAI (Microsoft AI) model family.

Let’s start with a key distinction: MAI and Copilot are not the same thing. Copilot is Microsoft’s chatbot, which runs primarily on OpenAI’s technology. MAI models, by contrast, are built on Microsoft’s own in-house large language models (LLMs). At Build 2026, Microsoft announced four primary model series, not counting the coding-specific MAI-Code:

• MAI-Thinking-1, a reasoning modelMicrosoft is calling these models experimental and describes them as being in a “limited preview” state, so don’t expect a finished product. But if you’re curious, you can access them for free via Microsoft’s Playground site. Just be aware that MAI-Thinking-1 is still “coming soon” and only available in early access to select users, so most people can’t try it yet.Below are my first impressions of each model. I’m speaking strictly from a consumer perspective, so these models may still have value in enterprise settings.

MAI-Thinking-1 Is Decent, But Why Pick It Over the Competition?

MAI-Thinking-1 is Microsoft’s first reasoning model, designed to handle complex prompts and topics. Microsoft compares it to Claude’s Sonnet model, claiming users preferred it in a blind side-by-side test conducted by Surge. For context, Anthropic markets Opus as Claude’s high-end reasoning model and Sonnet as the best blend of speed and intelligence.

In my testing, Sonnet (even on its medium intelligence setting) outperformed MAI-Thinking-1. For one, Microsoft’s model can’t access the internet, which is a deal-breaker for many tasks. I also didn’t notice any meaningful improvements in accuracy, response quality, or speed when I asked about the nuances of Path of Exile 2 game mechanics or help setting up a database structure. MAI-Thinking-1 isn’t unintelligent, but its limitations and overall performance don’t make a strong case for choosing it over competitors.

MAI-Image-2.5 Is Better, But Not a Game-Changer

MAI-Image has improved significantly since its first release in October 2025, when it lagged far behind top image-generation models. Today, MAI-Image-2.5 is much better, but it still doesn’t match Gemini’s Nano Banana Pro, one of our Technical Excellence winners.

I generated images of a suburban home, a comic, and a diagram using both models. Nano Banana Pro consistently produced sharper images, while MAI-Image-2.5 struggled, especially with text. The comic and diagram showed distorted text, a problem Nano Banana Pro didn’t have. If MAI-Image-2.5 is your only option, it can get the job done, but I wouldn’t recommend it as your go-to AI image generator.

MAI-Transcribe-1.5 Works Well Enough, But Doesn’t Shine

MAI-Transcribe-1.5 converts audio files to text, and it’s free to use. You just upload an audio file and get results in seconds. But it’s not better than other transcription tools. I used a GoTranscript transcription test to compare it with Gemini. MAI-Transcribe-1.5 made 13 mistakes, while Gemini made only six. I then gave both models a hardcore song to see if they could parse the lyrics. Both made errors, but MAI-Transcribe-1.5’s transcription cut off before the song ended.

Gemini isn’t even marketed as a transcription tool, which makes this comparison especially telling. MAI-Transcribe-1.5 is fine for quick jobs, but it’s far from best-in-class.

MAI-Voice-2 Sounds Robotic, and That’s a Problem

AI voices are everywhere, from AI music to Copilot conversations. Most still sit in the uncanny valley, not as fake as old digital voices but not quite human. Some, like Sesame’s tech, are genuinely impressive. MAI-Voice-2 is not one of them.

Its Playground interface is clean and easy to use, and you can choose from multiple languages and styles. But no matter what you pick, the voice sounds robotic. The combination of audio quality, breathiness, cadence, and intonation just doesn’t feel human. I’ve heard worse, but Microsoft’s dedicated voice model is far from the best.

Like Copilot, the MAI Models Are Mostly Unremarkable

Microsoft’s new MAI models are fine, and that’s about the best I can say. I feel the same way about Copilot, which stands out more for its deep integration with Microsoft’s ecosystem than for its underlying model. I understand why Microsoft wanted to showcase its in-house models at Build, but my initial testing suggests they aren’t ready for the spotlight. That said, the rapid improvement of MAI-Image over the past few months gives me hope. I’ll keep testing them, and I expect they’ll get better quickly.