Thank you. Good morning everybody. You know, we really are living in the most remarkable times. Since I started working in AI, the compute that we use to train frontier models has increased by a trillion fold. That's 12 orders of magnitude of computation in just 15 years. It's now clear that a consistent exponential increase in computation leads to predictable advances in AI capabilities. And in the next few years, we're going to see three more orders of magnitude of compute applied to train frontier models. Intelligence is now a function of compute. Log linear hill climbing has become the norm. The scaling laws are clearly holding, and it is a remarkable time in our industry.
And so in this context, we at MAI are building towards what we call Humanist Superintelligence. State of the art AI capabilities that are explicitly designed to serve people and organizations and not replace them, because the type of AI that we create really does matter. We need an AI that places humanity first, that always prioritizes human well-being and human progress. This is the core philosophy and motivation behind our superintelligence efforts at Microsoft, and it shapes everything that we do. And as a platform company, our job and our commitment is to keep you developers building at the absolute frontier.
So today, we are very excited to announce a family of seven new models across image, voice, transcription, and coding. These are all built with real attention to detail and a commitment to making very practical and efficient tools that are tuned to just how you work in the real world.
So first up, MAI Image 2.5 and its Flash variant, two super strong models that deliver a step change in quality, now at number two on the leaderboard, surpassing the score of Nano Banana 2 on image editing. They give you precise editing with incredible control and consistency. Flash is here for super efficient production workloads, while 2.5 gives you that maximum fidelity and professional-grade performance. They are live in PowerPoint today, they're rolling out to OneDrive, and right now you can access them on Foundry at a market leading quality per dollar.
Next up we've got MAI Transcribe 1.5. This is the best transcription model in the world. State of the art accuracy across 43 languages, beating out Gemini and OpenAI's flagship transcription models. We've optimized it for real world use so that you can produce highly accurate transcripts for any bespoke use case, five times faster than all rival models. It's now being integrated inside of GitHub, Teams, Copilot, Dynamics 365 Contact Center, and it's now also available in Foundry, where I'm very excited to say it is the fastest, most efficient, and most cost effective transcription model of any of the hyperscalers out there.
So paired with that, we've got MAI Voice 2. This is our latest speech generation model. It has beautiful prosody, natural sounding delivery, fine-grained emotional control. And it's available in 15 languages, with many more coming soon. We're also announcing Voice 2 Flash, and that provides the very best value and speed for ultra latency sensitive voice agents, which of course is the big thing in 2026.
Next up, our text foundation model. MAI Thinking 1. This is our first reasoning model, and it's exceptionally strong in our target use cases of reasoning and SWE tasks. It's a 35 billion active parameter MOE with a 256k context window. That means that it competes in the medium size weight class, where it's certainly punching above its weight, and independent human raters on Surge prefer it in overall quality side by side versus Sonnet 4.6. It's achieved 97% on AME 2025. Obviously, the key measure of its general purpose reasoning abilities. But most importantly of all, it's now at 53% on SWE Bench Pro, which places it right alongside Opus 46 at least on the toughest coding benchmark that's out there. So we're very happy with that.
Now, there's plenty more for us to do as we get this into production and hill climb against real-world tasks and real-world traffic. What is actually most remarkable about this model, we think, is that it has climbed entirely from the bottom, and that means that it hasn't targeted any of the benchmarks specifically. And it's done so with absolutely zero distillation. And to us, this is critical because it means that the model is created with an enterprise grade, clean and commercially licensed data lineage. That means that you can put it into production in a very trustworthy way with complete confidence.
Now finally, I'm incredibly excited to announce MAI Code 1 Flash. This is our new inference efficient coding model, which has been especially tuned for VS Code and of course GitHub Copilot CLI. It achieves 51% on SWE Bench Pro despite having just 5 billion parameters. And so it's much closer to Haiku in terms of size, but cheaper in cost, delivering really strong coding performance at great inference efficiency. And it's rolling out today inside of VS Code, alongside distribution on Foundry and optimization for our 1P products. We're also very excited to make our models available on OpenRouter as well as Fireworks and Baseten. So this means that for the first time, you're going to be able to tune the weights directly yourself in an ecosystem of your choice.
Now, across this entire family, safety and security have been built in from the start. Our voice models come with protections against unauthorized cloning. Everything is watermarked from scratch. We've reduced our over-refusals, improved representation, including for people with disabilities. We're also publishing a very detailed technical report today to give you a full and transparent understanding of how we put all of this together.
Now, one of the things I'm particularly excited about is that we have been carefully co-designing our models with our own silicon. So that means that we've optimized MAI Thinking 1 on our very own Maia 200 chip and benchmarked it head to head against the GB-200. And so on top of the 30% performance improvement that Satya talked about earlier, we're now seeing a further 1.4x performance per watt gain when we run our MAI models on the Maia 200, end to end. And that's huge, because as everybody knows, at this scale, every watt counts and silicon and model co-design is a really key advantage that we think is going to help keep everybody here right on the frontier with the most efficient and most powerful thinking and coding agents out there. We're also super excited that these faster and more efficient MAI models are coming to the N1X that Satya mentioned a few moments ago, and we think that's going to be able to deliver the very best performance on Windows in a few months time.
Now to us, this is what owning the full stack end-to-end looks like. It's the foundation of Microsoft Frontier Tuning, lets you customize the MAI models using our full stack hill climbing machine right where you want it, and it means that the disciplined and very relentless engineering that has gone into building our models is now available to all of you, on a platform that you can trust, working on your behalf to create custom agents that you will control.
So the really big thing, of course, that's happened in the last year is these RLEs, reinforcement learning environments, these unique training gyms for your AIs. They create company and task specific agents, adapted only to you, built on MAI models. So for example, within Microsoft we use our RLEs combined with our MAI models to climb towards the best agentic use cases on Excel. Our MAI tuned model is now on par with GPT 5.4 on public and private benchmarks, whilst at the same time being ten times more efficient on cost.
Thank you. You know, and many other early adopters are seeing similar results. When we've tuned our models on McKinsey's tasks, MAI delivered the highest win rate, even outperforming GPT 5.5 and again rate, delivering 10x greater efficiency on cost. So to us, this is the advantage of very carefully calibrated Frontier Tuning. And importantly, unlike with some of the other companies, with MAI you don't rent intelligence from a shared model that learns from everybody. Only you keep the benefits of your hard-earned workflows, know-how, knowledge and your own institutional data. Only you get to control the resulting model. And so with us, the RLEs and the models that you build inside of them, they become your moat. I really think this is distinct. It marks a new era in AI that we're all very, very excited about.
Okay, so now just one final announcement that I'm very excited about. We are taking customization and co-creation of our models to the highest level possible. On what I think of as perhaps the most important application of AI, healthcare. So today, we're very proud to be announcing that we're partnering with Mayo Clinic to jointly develop a new frontier model for health and then deploy it around the world in their hospitals and beyond. So this morning, please help me in welcoming to the stage a physician, groundbreaking researcher, president and CEO of the Mayo Clinic, Doctor Gianrico Farrugia.
Thank you so much for being here, Gianrico. Now, of course, everyone will recognize Mayo as perhaps the leading hospital in the world with an incredible track record of research and innovation and clinical practice. Tell us a little bit more about what you hope to get out of our collaboration.
