Microsoft just released its first proper reasoning model, and it’s not from OpenAI or Anthropic.
It’s called MAI‑Thinking‑1, and it landed in public preview inside Microsoft Foundry this month.
If you’re running a Microsoft‑first business, this is worth pausing on, not because it’s flashy, but because it quietly changes who’s actually building the intelligence behind your Copilot experience.
For years, Microsoft’s AI story has really been OpenAI’s story with a Microsoft badge on it. GPT models power most of what you see in Copilot, and Azure has been the place you go to rent access to someone else’s model.
MAI‑Thinking‑1 is Microsoft saying it wants to own more of that stack itself and for business leaders trying to plan sensibly around AI, that shift matters more than another benchmark score.
What Microsoft Actually Released
MAI‑Thinking‑1 is a text based reasoning model, built by Microsoft’s own AI team rather than licensed or distilled from another lab.
It’s what’s called a sparse Mixture of Experts model, which sounds technical but the idea is simple: instead of one giant model doing everything, it’s made up of many smaller specialist components that get called on depending on the task.
That design gives it roughly one trillion parameters in total, but only around 35 billion are active for any single request, which keeps it faster and cheaper to run than a model that size would normally be.
It also handles a genuinely long context window, up to 256,000 tokens, which Microsoft says is enough to process a document around 600 pages long in one go. For a business that deals with lengthy contracts, policy manuals or years of case notes, that’s a real capability, not a marketing line.
On performance, Microsoft reports MAI‑Thinking‑1 scores 97.0% on the AIME 2025 maths benchmark and 94.5% on AIME 2026, and says it’s roughly level with Anthropic’s Claude Opus 4.6 on a tough coding benchmark called SWE‑Bench Pro.
Microsoft also ran blind side by side testing with professional raters through a partner called Surge, across 1,276 tasks, and says people preferred MAI‑Thinking‑1’s answers over Claude Sonnet 4.6.
Worth remembering these are Microsoft’s own reported numbers rather than independently verified results, so treat them as a solid signal rather than gospel.
Why Microsoft Didn’t Just Use GPT Again
Here’s the interesting part. Microsoft says it deliberately built MAI‑Thinking‑1 without distilling from any third party model. In plain terms, that means it didn’t take a shortcut by learning from another company’s AI outputs.
It trained the model from scratch on what it describes as clean, traceable, enterprise grade data.
Why does that matter to you as a business owner? Because it’s about accountability.
If Microsoft can trace exactly what data shaped the model, it’s in a stronger position to explain and defend how the model behaves, which is a genuinely useful thing when you’re trying to satisfy a board, a regulator or a nervous client about how your AI tools actually work.
It won’t remove every governance question you have, but it does mean Microsoft can answer more of them with a straight face.
Where This Actually Shows Up for You
MAI‑Thinking‑1 sits inside Microsoft Foundry, which is Microsoft’s platform for building and deploying AI models and agents.
It’s listed as a “Direct from Azure” model, meaning it’s billed, supported and governed entirely within Microsoft’s own environment rather than passing your data or your invoice through another vendor.
It uses a widely supported API style called Chat Completions, and it supports function calling, so it can be wired into agents that need to call tools or pull information from other systems.
Practically, that means if you’re a Foundry customer building an internal agent, or you’re working with a partner like CG TECH on a Copilot Cowork rollout, MAI‑Thinking‑1 becomes another option in the model picker.
It’s not going to replace GPT‑5.6 or Claude Opus 5 for every task, and Microsoft has been clear that this is public preview, not general availability, so I wouldn’t be putting it straight into a production process handling anything sensitive just yet. But it’s a real, usable option starting now.
The Bit Business Leaders Should Actually Care About
I keep coming back to the same point when businesses ask me about new AI releases: the model is only ever half the story.
I made this case a few weeks ago when I wrote about why your knowledge layer matters more than your model, and MAI‑Thinking‑1 doesn’t change that. A brilliant reasoning model working off messy SharePoint permissions and duplicate documents will still give you messy, unreliable answers.
What MAI‑Thinking‑1 does change is your options for the engine sitting on top of that knowledge layer, and possibly your cost line, since Microsoft is explicitly pitching this as cost efficient reasoning for high volume enterprise tasks.
There’s also a governance angle here that connects to something I wrote about recently on open weight AI versus hosted AI.
You’ve now effectively got three lanes running through Foundry: Microsoft’s own MAI models, hosted third party models like GPT‑5.6 and Claude, and open weight models you deploy and manage yourself.
Each lane has a different risk and cost profile. MAI‑Thinking‑1 adds a genuinely useful middle option, one where Microsoft owns the whole chain from training data through to hosting, which some businesses will find easier to govern than a model built by an external lab.
Where I’d Be Cautious
I don’t think every business needs to rush out and test this model this week. A few things are worth sitting with first.
It’s still in public preview, which means behaviour, pricing and availability can shift before it’s generally available. Microsoft has also been upfront that MAI‑Thinking‑1 hasn’t been evaluated for fully autonomous use on untrusted external content without human oversight, and it shouldn’t be the sole decision maker in high stakes areas like credit, employment or safety decisions.
That’s a sensible warning, and one worth building into any pilot you run.
If you’re already deep into Copilot, and you’ve read my earlier piece on what actually changes with the new unified Copilot app, you’ll know Microsoft has been busy simplifying its AI product line generally this month.
MAI‑Thinking‑1 fits that same pattern: Microsoft consolidating and taking more ownership of its AI stack, rather than throwing new features at the wall.
What I’d Actually Do Next
If you’re running IT or digital strategy for a Microsoft‑first business, here’s where I’d start.
First, don’t treat this as an urgent switch. Treat it as an option worth testing quietly, in parallel with whatever you’re already using.
Second, pick a handful of real tasks, maybe 20 to 30, where reasoning over long documents matters, things like contract review, policy summarisation or case history analysis. Run them through MAI‑Thinking‑1 in Foundry and compare the results against your current model on accuracy, speed and cost.
Third, loop your governance or compliance lead in early, not after the pilot’s finished. Ask specifically how this model’s data provenance and closed weights change your risk assessment compared with what you’re using now.
Fourth, keep an eye on the general availability announcement. Public preview pricing and behaviour can shift, and you don’t want to build a business case on numbers that change under you.
Microsoft building its own reasoning model isn’t a headline that needs to change your AI roadmap overnight. But it is a signal worth paying attention to, because it tells you where Microsoft is putting its energy, and it gives you one more genuinely useful option when you’re deciding which model does the thinking behind your next AI project.
That’s why it matters: not because MAI‑Thinking‑1 is the smartest model on the market, but because it’s a sign of where the ground is shifting under your existing AI stack.
If you want a hand working out whether MAI‑Thinking‑1 is worth testing for your business, or you just want a second opinion on your broader AI model strategy, get in touch with the team at CG TECH.
About the Author
Carlos Garcia is the Founder and Managing Director of CG TECH, where he leads enterprise digital transformation projects across Australia.
With deep experience in business process automation, Microsoft 365, and AI-powered workplace solutions, Carlos has helped businesses in government, healthcare, and enterprise sectors streamline workflows and improve efficiency.
He holds Microsoft certifications in Power Platform and Azure and regularly shares practical guidance on Copilot readiness, data strategy, and AI adoption.
Sources
- Introducing MAI-Thinking-1 — Microsoft AI
- New MAI models in Microsoft Foundry across text, image, voice and search — Microsoft Tech Community
- MAI-Thinking-1: 1T MoE, No Open Weights — ExplainX
- Microsoft Puts MAI-Thinking-1 in Foundry Public Preview — Magica
- Microsoft Build 2026: MAI keynote transcript — Microsoft AI