On 2 June, Microsoft's AI chief Mustafa Suleyman walked on stage at Build and announced seven new AI models in a fast fifteen-minute keynote. Image, voice, transcription, reasoning, coding. It was a lot to absorb, and if you run a team rather than a model lab, the natural response is to feel slightly behind before you've finished your coffee. The number of models is not the story. What they signal about the next twelve months inside your organisation is.
The short version
Microsoft has stopped being mainly a reseller of other people's AI and started shipping its own. The new models are capable, cheap to run, and already turning up inside the software your team uses every day. For most organisations the headline isn't any single model. It's that genuinely useful AI is about to arrive by default, whether or not you have a plan for it.
What Microsoft's new AI models actually do
The seven models, in plain English, grouped by what they do:
- Reasoning: MAI-Thinking-1. Microsoft's first reasoning model. On the company's own figures it lands alongside a leading frontier model on the hardest coding benchmark it cited, and scores near the top on a tough maths test, at a medium size rather than a giant one. Microsoft made a point of saying it was trained without copying from other labs, on properly licensed data, which is a deliberate pitch to nervous enterprise buyers.
- Coding: MAI-Code-1-Flash. A tiny, cheap coding model tuned for VS Code and GitHub Copilot. The interesting part is the size, around five billion parameters, doing work that recently needed something far larger.
- Image: MAI-Image-2.5, plus a faster Flash version. Strong image generation and editing, already live in PowerPoint and rolling into other Microsoft tools.
- Transcription and voice: MAI-Transcribe-1.5 and MAI-Voice-2. Transcription across 43 languages that Microsoft claims is the fastest and most accurate going, plus natural-sounding speech generation and a low-latency version built for voice agents.
One habit worth keeping: every benchmark here is the vendor's own. Encouraging, not gospel.
Why this matters for your organisation
Strip away the model names and three shifts actually matter to a decision-maker.
Microsoft is now a model maker, not just a middleman. For years its AI story was largely OpenAI's technology wrapped in Microsoft products. It now has a competitive stack of its own. More serious competition at the frontier tends to mean lower prices and faster progress for everyone buying, and less of your future riding on a single supplier.
This arrives by default, not by procurement. These models are going live inside PowerPoint, Teams, Copilot, VS Code and Dynamics. That changes the nature of the decision in front of you. AI in your organisation is becoming a default setting rather than a project you choose to start. You want a position on governance, data and disclosure before the features switch on, not a scramble afterwards.
The real pitch to leaders is ownership. The line Microsoft leant on hardest was "Frontier Tuning": build agents on models adapted to your own workflows and data, where the result is yours and your institutional knowledge stays with you rather than feeding a shared model everyone else also rents. That is a genuinely attractive idea, and it comes with an obvious catch. It deepens your reliance on Microsoft's stack, and it only works if your processes and data are in good enough shape to train on in the first place. Most organisations badly overestimate how ready theirs are.
Underneath all of it, the economics moved. A five-billion-parameter coding model and a medium reasoning model getting close to frontier results, at a fraction of the cost, is what makes rolling AI out across a whole organisation affordable. Cheap models that are good enough, running everywhere, will change far more workplaces than one brilliant, expensive model running in a single corner of the business.
What I'd actually do about it
Here is the trap. The instinct after an announcement like this is to chase the models. Don't. The capability ceiling rose again this week, and it will rise again next month. Chasing it is a treadmill, and it is the wrong race.
The thing that quietly costs you money is the distance between what these tools can now do and what your people actually do well with them. That distance is an organisational problem, not a technical one, and no new model closes it for you. I've written before about the competencies that decide whether AI capability compounds across a team or stays stuck with a few keen individuals, and no model launch, this week's or next month's, changes that.
So, three practical moves:
- Take a stance on what's already arriving. The AI features landing in your Microsoft tools need a clear line on data, governance and disclosure. Decide it now, while you still can, rather than after someone has pasted something sensitive into a transcript tool.
- Get good at the basics on one or two real workflows. Briefing AI clearly, judging what it gives back, and improving the next attempt is a learnable skill, and it matters far more day to day than which model sits underneath. That is exactly what our AI Fluency course walks through.
- Treat Frontier Tuning as a question, not a purchase. Before anyone sells you a custom model, answer the unglamorous bit: what would you train it on, and is that data actually ready? For most teams, the honest answer is "not yet", and that is the real first project.
Microsoft framed the whole event around "humanist superintelligence", AI built to serve people rather than replace them. It's the right ambition, and I genuinely hope they mean it. But a slogan doesn't make adoption humane. Design does. Whether this wave of cheaper, better, ever-present AI makes your organisation more capable or just more anxious is decided by how you build it in, not by which model you pick.
Common questions about Microsoft's new AI models
Do Microsoft's new AI models replace ChatGPT or OpenAI for businesses? Not as a switch you have to make. Microsoft still partners with OpenAI, but it now also has its own competitive models and is building them into its own products like PowerPoint, Teams and GitHub Copilot. For most organisations this means more choice and less dependence on a single AI provider, rather than a migration you need to plan.
Can you trust Microsoft's benchmark claims for the new MAI models? Treat them as the vendor's own figures. The numbers Microsoft shared, such as its reasoning model scoring alongside a leading frontier model on a hard coding benchmark, are promising, but every lab presents its work in the best possible light. What matters is how a model performs on your actual tasks, so pilot it on real work before you believe a leaderboard.
What is Microsoft's Frontier Tuning and should we use it? Frontier Tuning is Microsoft's name for adapting its models to your specific workflows and data so the resulting model is yours and your data stays with you. It is a real opportunity for competitive advantage, but only if your processes and data are in good enough shape to train on, and it deepens your reliance on Microsoft's stack. Treat it as a question to investigate, not a box to tick.
We're a smaller organisation. Does this announcement matter to us? Arguably more than it does to large enterprises. The cheaper, smaller models are exactly what make capable AI affordable at your scale, and they are arriving inside tools you already pay for. The advantage goes to whoever builds the habits to use them well, which is far less about budget than it is about capability.
The bottom line
None of this needs a panicked response. It needs a deliberate one. The organisations that get value from the next year of AI won't be the ones that adopted fastest. They'll be the ones that built the capability to use it well. If you want help turning this week's noise into a plan your team can act on, starting with the unglamorous question of whether your data is even ready, that's exactly what I help teams with. Get in touch.