Microsoft AI simply launched MAI-Considering-1, their first reasoning mannequin – and it’s already within the Microsoft Foundry mannequin catalog in public preview. Introduced August 12, 2026, that is MAI’s personal reasoning mannequin, skilled from the bottom up, not distilled from anybody else’s fashions.
And sure – I after all examined it out. I already deployed it to my very own Microsoft Foundry, so let’s take a more in-depth have a look at what it’s and why it issues for Future Work.
MAI-Considering-1 is a reasoning mannequin: you give it a immediate, and it produces an inside chain of thought earlier than it solutions. The fascinating half is that it allocates the reasoning effort adaptively based mostly on how advanced your immediate is. Easy query, much less pondering. Onerous query, extra pondering. That’s precisely the type of conduct you need when you find yourself paying per token.
Below the hood it’s a sparse Combination-of-Consultants (MoE) Transformer with 35B lively and ~1T complete parameters. MoE means solely the components of the mannequin wanted for every request get activated – functionality scales up with out compute scaling linearly with it. Microsoft calls it a medium-sized mannequin that stands among the many strongest in its weight class.
Some key numbers:
- 256K context window – Microsoft notes that is sufficient to match a 600-page doc
- Operate calling and assist for developer directions
- Constructed on the broadly used Chat Completions API, so migration is supposed to be simple
- Coaching cut-off: July 2026
- Textual content-only – no picture, audio or video in or out
Work is altering. And the fashions we construct our brokers on are altering even sooner.
Microsoft isn’t claiming that is the largest mannequin on the planet. They’re claiming it’s the finest in its weight class, and the numbers they revealed again that framing up:
- 97.0% on AIME 2025 and 94.5% on AIME 2026 – critical mathematical reasoning
- Toe-to-toe with Claude Opus 4.6 on SWE-Bench Professional for agentic coding
- Most well-liked over Claude Sonnet 4.6 in a blind human side-by-side analysis run with their associate Surge, spanning 1,276 duties in single-turn and multi-turn conversations
That final one is the one I discover most fascinating. Benchmarks measure functionality, however that human choice analysis measures whether or not the mannequin truly understands the duty, follows the directions, makes use of the appropriate degree of element and respects your time. That’s the stuff that decides whether or not folks maintain utilizing an agent you constructed – or quietly return to doing it by hand.
That is the half I feel enterprises ought to learn twice. Microsoft skilled MAI-Considering-1 with out distillation from third-party fashions, on knowledge they describe as clear, traceable and enterprise-grade, appropriately licensed.
Why does that matter to you and me? As a result of provenance is turning into a governance query, not only a analysis one. When your authorized, compliance or safety folks ask “what formed this mannequin?”, “we are able to account for it” is a really completely different reply than “it realized from one other mannequin that realized from one thing else”. High quality, provenance, management – that’s the pitch, and it’s a good one.
Microsoft frames this as a part of their broader work in direction of Humanist Superintelligence: AI designed to serve folks and organizations, not change them. Additionally they make some extent I actually like – {that a} mannequin refusing respectable requests beneath the guise of security is a defect too, not a function. They prepare unsafe compliance and pointless refusal as defects in the identical reward system. Anybody who has ever been refused by an AI for asking one thing fully regular is aware of precisely why that issues!
MAI-Considering-1 is priced at $2 USD per 1M enter tokens and $8 USD per 1M output tokens.
That’s the entire guess, actually: sturdy reasoning at a price-performance level that makes high-volume, always-on AI workloads economically viable. As a result of right here is the factor about brokers – the pilot is rarely the costly half. The costly half is when it really works, everybody begins utilizing it, and it runs all day on daily basis throughout the group. A reasoning mannequin that’s inexpensive sufficient to depart operating is a special proposition than one you solely deliver out for particular events.
Mix that with adaptive reasoning effort and also you get a mannequin that doesn’t burn tokens pondering laborious about simple questions. Sensible, like actually sensible.
MAI-Considering-1 is a Direct from Azure mannequin in Microsoft Foundry. In follow:
- Secured and managed by Microsoft – single license, constant assist, no third-party dependencies
- Unified billing and governance, with PTU portability throughout fashions hosted on Azure
- Pay-as-you-go flexibility, or reserve PTUs if you need predictable efficiency and financial savings
- Take a look at, deploy and swap between fashions inside one platform
It additionally plugs into Foundry’s built-in analysis, observability, security and deployment capabilities. In case you are constructing brokers for actual manufacturing use, that surrounding toolset issues not less than as a lot because the mannequin itself.
MAI-Considering-1 is obtainable on varied different areas additionally, I couldn’t screenshot them suddenly so I chosen some areas. The purpose: it’s fairly broadly out there, however there are some areas that don’t have it.
Microsoft calls out three use case areas, and all three are very recognizable from actual buyer work:
- Enterprise deployments – 256K context, clear knowledge provenance, perform calling, advanced instruction following
- Coding workflows – studying code, enhancing recordsdata, operating exams, bug fixing, observing failures and recovering from intermediate errors
- Advanced reasoning – particularly quantitative work like monetary modeling, statistical evaluation, market sizing and forecasting
That 256K context window is the one I maintain coming again to (though it’s not 1M like with Claude..). Lengthy agent traces with out chunking and stitching them collectively is a real quality-of-life enchancment when you find yourself constructing multi-step brokers. Anybody who has constructed a workflow that needed to summarize its personal historical past to outlive is aware of the ache.
It’s public preview, so deal with it accordingly – that is for testing and constructing, not for betting your manufacturing workload on as we speak.
A number of extra issues price figuring out earlier than you deploy:
- It’s not designed as an autonomous decision-maker in consequential domains – authorized, monetary, medical, employment, instructional, housing, credit score, safety-critical. Not an alternative choice to skilled recommendation in regulated fields.
- It has no native instrument interface. Device use is mediated completely by your utility, which suggests you personal the safety boundary round something you expose to it. Please learn that sentence once more in case you are constructing brokers.
- It’s not evaluated for totally autonomous agentic deployments appearing on untrusted exterior content material with out human oversight or harness-level controls.
- Language protection varies. It’s primarily optimized for English, and likewise helps German, Spanish, French, Italian, Portuguese, Chinese language (Simplified), Russian, Hindi, Japanese, Korean, Arabic and Hungarian, with extra restricted protection for Hebrew, Turkish, Persian, Thai, Vietnamese, Indonesian and Ukrainian amongst others.
Strive MAI-Considering-1 in Microsoft Foundry Public Preview API and Playground can be found. Deployment was genuinely simple – discover it within the catalog, deploy, and you might be testing in minutes. Setup couldn’t be a lot easier.
MAI-Considering-1 isn’t making an attempt to be the largest mannequin within the room. It’s making an attempt to be the one you may truly afford to run all day – and that may be a rather more helpful ambition for the type of work most of us are doing proper now. Sturdy reasoning, actual agentic coding chops, a 256K window, clear and traceable knowledge, and $2/$8 per million tokens.
For me probably the most thrilling sign isn’t the benchmark desk in any respect. It’s that Microsoft is constructing its personal reasoning functionality from the bottom up, on knowledge it could actually account for, inside the identical Foundry platform the place the analysis, observability and governance already dwell. That could be a very completely different basis to construct enterprise brokers on than we had even a yr in the past.
I’m additionally ready and hoping that this mannequin will come to Copilot Cowork quickly, as it should assist to decrease prices utilizing Cowork!
Have you ever already deployed MAI-Considering-1 in your personal Foundry? I want to hear what you might be constructing with it – drop a remark and let me know!
Sources: Introducing MAI-Thinking-1 | Microsoft AI · MAI-Thinking-1 in the Microsoft Foundry model catalog
Revealed by
I work, weblog and discuss Future Work : AI, Microsoft 365, Copilot, Loop, Azure, and different providers & platforms within the cloud connecting digital and bodily and folks collectively.
I’ve 30 years of expertise in IT enterprise on a number of industries, domains, and roles.
View all posts by Vesa Nopanen
You might also like
More from Metaverse Global
NEAR Governance Proposal Targets Lower Token Issuance and Dilution
Key Highlights NEAR proposes slicing token issuance from 2.5% to 1.6% over 24 months. The decrease issuance might stop about 66 …
Hawley–Murphy AI Bill Targets Agent Hacking Liability as Crypto Risks Emerge
Senators Josh Hawley (R-Mo.) and Chris Murphy (D-Conn.) are making ready to introduce the AI Agent Accountability Act, which …
