Microsoft’s 7 New AI Models: What They Are and Why They Matter
Microsoft’s MAI Models Explained: What They Are, Why They Matter, and What Changes Now
At Build 2026, Microsoft unveiled seven AI models it built from scratch — no OpenAI, no distillation. Here’s a plain-English breakdown of each model, the strategic shift behind them, and what it means for developers and everyday users.
Microsoft has spent the past four years as the most powerful distributor of someone else’s AI. Its $13 billion investment in OpenAI gave it access to GPT-4, GPT-4o, and the o-series models — which it embedded into Copilot, Azure, Office, and GitHub. The strategy worked remarkably well. Microsoft became the first tech company to deeply integrate generative AI into productivity software at scale. But it also meant that Microsoft’s AI roadmap was, to a significant degree, OpenAI’s roadmap.
At Build 2026 on June 2 in San Francisco, Microsoft AI CEO Mustafa Suleiman announced something different. Seven new AI models — built entirely from scratch by Microsoft’s own Superintelligence team, with zero distillation from OpenAI or any other external model. The family is called MAI, and it covers reasoning, coding, image generation, voice synthesis, and speech transcription. Microsoft’s stated goal: “long-term self-sufficiency.”
This article explains what each model does, how they compare to the competition, and what the shift means — for Microsoft, for OpenAI, and for the developers and businesses that use both.
Microsoft unveiled seven MAI models at Build 2026 on June 2, 2026: MAI-Thinking-1 (reasoning), MAI-Code-1 and MAI-Code-1-Flash (coding), MAI-Image-2.5 and MAI-Image-2.5-Flash (image generation/editing), MAI-Transcribe-1.5 (speech-to-text, 43 languages), and MAI-Voice-2 (voice synthesis, 15+ new languages). All built from scratch by Microsoft’s Superintelligence team with no external distillation. MAI-Thinking-1 is a 35-billion-active-parameter Mixture of Experts model (~1 trillion total parameters) with a 256K context window that scores 53% on SWE-Bench Pro — matching Claude Opus 4.6 — and is preferred over Claude Sonnet 4.6 in blind tests. MAI-Code-1-Flash (5B parameters) is already the default in GitHub Copilot. Microsoft AI CEO Mustafa Suleiman acknowledged Anthropic “still holds a lead of a few months” — an unusually honest admission at a product launch.
Microsoft Build 2026 official keynote — MAI model announcements begin around the 45-minute mark. Source: Microsoft on YouTube. All rights Microsoft.
Why Microsoft Is Building Its Own Models Now
To understand why this matters, it helps to understand what Microsoft’s AI situation looked like before Build 2026. The company had invested $13 billion in OpenAI and received the right to use OpenAI’s models commercially through Azure. It also had its own Phi series of small language models, designed for on-device and edge deployment. But for enterprise and developer use — the serious, production-grade AI that runs inside Copilot and GitHub — Microsoft was almost entirely dependent on OpenAI.
That dependency created three problems. First, cost — Microsoft pays OpenAI for every API call, which means every Copilot interaction has a margin structure determined by OpenAI’s pricing. Building its own models allows Microsoft to serve the same use cases at lower token cost. Second, control — if OpenAI’s models change, get deprecated, or become unavailable for any reason, Microsoft’s products break. Third, competition — OpenAI is increasingly a direct competitor to Microsoft, offering its own enterprise products that bypass Azure entirely.
Imagine a restaurant that sources all its ingredients from one supplier. The food is excellent, but the restaurant has no control over prices, availability, or what the supplier does next — including whether the supplier decides to open its own competing restaurant. Microsoft has been in that position with OpenAI for four years. Building the MAI models is the equivalent of opening its own kitchen garden. It doesn’t mean Microsoft stops using OpenAI — but it means it no longer has to.
The Seven MAI Models — What Each One Does
Microsoft’s first in-house reasoning model. 35 billion active parameters in a sparse Mixture of Experts (MoE) architecture — approximately 1 trillion total parameters. 256K token context window (enough to process a 600-page document in a single pass). Built for complex multi-step instructions, long-context reasoning, and code generation. Trained entirely from scratch on commercially licensed data with zero distillation. Scores 97% on AIME 2025 and 94.5% on AIME 2026. Achieves 53% on SWE-Bench Pro — matching Claude Opus 4.6. Independent raters (via Surge) preferred it over Claude Sonnet 4.6 in blind tests. Available on Azure AI Foundry in private preview. Mustafa Suleiman acknowledged Anthropic still holds “a lead of a few months” at announcement.
Microsoft’s first dedicated coding model, built specifically for GitHub Copilot and VS Code. MAI-Code-1-Flash is a 5-billion-parameter model already live as the default in GitHub Copilot’s auto-picker and VS Code. Outperforms Claude Haiku 4.5 across all four core coding benchmarks tested — including a 16-point lead on SWE-Bench Pro (51.2% vs 35.2%) and uses up to 60% fewer tokens than comparable models on the same tasks. Also available on Fireworks AI, Baseten, and OpenRouter for third-party developers.
Microsoft’s first model capable of both text-to-image generation and image-to-image editing. Ranks #3 on the Arena AI leaderboard for text-to-image and #2 for image-to-image workloads. The Flash variant is optimised for faster, lower-cost image tasks. Replaces Microsoft’s previous reliance on third-party image models in Designer and other products.
Speech transcription model supporting 43 languages. Designed for Teams meeting transcription, Copilot voice features, and enterprise dictation workflows. Replaces Azure’s previous third-party transcription dependencies for high-volume, multilingual workloads.
Text-to-speech model adding new synthetic voices in 15+ additional languages beyond what Microsoft’s previous voice services covered. Powers Copilot voice responses, Teams meeting narration, and accessibility features across Microsoft’s product suite.
Microsoft states that MAI-Thinking-1 “matches Opus 4.6 on coding abilities on SWE-Bench Pro” and that “independent raters prefer it to Sonnet 4.6 in blind tests.” These are Microsoft’s own claims from its official blog post. Independent third-party evaluations have not yet been published at time of writing. Benchmark comparisons are a standard part of AI model launches — but they should be read as directional indicators, not definitive performance verdicts. External validation typically follows weeks after announcement. The claims are plausible given the 35B parameter scale and training approach, but should be treated as unverified until independent testing confirms them.
What Else Microsoft Announced at Build 2026
The MAI models were the headline, but Build 2026 was a broader statement about Microsoft’s direction. Several other announcements are worth understanding in context.
Microsoft Scout
An always-on personal AI agent that works proactively across Teams and Outlook — scheduling meetings, preparing agendas, and summarising email threads before you ask. Unlike Copilot, which responds to prompts, Scout runs continuously in the background and surfaces relevant information ahead of your next task. It connects to your calendar, email, and meeting history and acts more like a chief of staff than a search assistant.
Surface RTX Spark Dev Box
A developer workstation powered by Nvidia’s RTX Spark chip, capable of running models up to 120 billion parameters locally — without cloud connectivity. 1 petaflop of AI compute and 128GB of unified memory. Designed for developers who need frontier-class AI inference on-device for privacy, latency, or cost reasons. This connects directly to the trend of AI moving off the cloud and onto local hardware.
Microsoft Execution Containers (MXC)
A Windows-level security sandbox that isolates AI agent actions from the underlying operating system. When an AI agent performs a task — browsing the web, editing a file, running code — MXC confines it to a sandboxed environment so a compromised or misbehaving agent cannot affect the rest of the system. This is a direct response to enterprise IT concerns about agentic AI security.
Majorana 2 Quantum Processor
Microsoft’s second-generation quantum chip, capable of 1-microsecond qubit operations. Designed for scalable quantum computing using topological qubits. Not consumer-relevant today — but the announcement signals Microsoft is building the long-term compute infrastructure that could underpin post-silicon AI development in the 2030s.
Mayo Clinic Healthcare AI Partnership
Microsoft and Mayo Clinic announced a collaboration to build a frontier AI model for healthcare diagnostics, combining Microsoft’s reasoning capabilities and compute with Mayo Clinic’s clinical expertise and patient data. This is the first time a major health institution has partnered with a technology company to build a genuinely frontier-class medical AI model rather than deploying an existing one.
What This Means for OpenAI
The relationship between Microsoft and OpenAI has always been unusual. Microsoft invested $13 billion and received commercial rights to OpenAI’s models. OpenAI received the capital and cloud infrastructure it needed to scale. Both benefited enormously. But the interests were never fully aligned.
OpenAI has been building its own enterprise sales motion — selling directly to businesses through ChatGPT Enterprise, bypassing Azure. Microsoft has been watching that happen while its own Copilot products compete for the same customers. The MAI models don’t end the Microsoft-OpenAI relationship — OpenAI’s models will still be available on Azure, and Microsoft will still use them where they’re the best option. But they fundamentally change the negotiating dynamic. Microsoft no longer needs OpenAI in the way it did before. That matters for contract terms, pricing, and the long-term structure of the relationship.
If you build on Azure or use GitHub Copilot, the MAI models give you a new set of options that are likely to be cheaper per token than equivalent OpenAI models for similar tasks — because Microsoft owns the entire stack. MAI-Code-1 is already live in Copilot and VS Code. MAI-Thinking-1 is in private preview on Azure AI Foundry. The models will also be available on Fireworks AI, Baseten, and OpenRouter for developers who don’t use Azure. The practical implication: more model choice at lower cost for standard enterprise tasks, without losing access to OpenAI models for cases where they’re genuinely better.
My Take
The MAI announcement is significant — but the framing of “Microsoft shocked the AI world” overstates it. The AI world largely expected this. Microsoft has been signalling for two years that it wanted to reduce its OpenAI dependency. The Phi series of small models was the beginning. MAI is the middle chapter.
What’s genuinely notable is the execution and the honesty. MAI-Thinking-1 achieving 53% on SWE-Bench Pro and 97% on AIME 2025 — if those hold under independent evaluation — is a real engineering achievement for a first-generation in-house model. More interesting is that Suleiman publicly acknowledged Anthropic “still holds a lead of a few months.” That’s not the language of a company claiming victory. It’s the language of a company that has closed a significant gap and knows exactly where it still stands. That kind of measured honesty is actually more credible than a triumphant launch claim would be.
The honest caveat is that we’re still at the announcement stage. Microsoft’s benchmark claims are plausible but self-reported. Independent evaluations, developer feedback from the private preview, and real-world performance data will determine whether MAI models become the default choice for Azure developers or remain a secondary option alongside OpenAI. That story will play out over the next six to twelve months.
Frequently Asked Questions
What are Microsoft’s MAI models?
MAI stands for Microsoft AI. The MAI models are a family of seven AI models built entirely from scratch by Microsoft’s Superintelligence team, announced at Microsoft Build 2026 on June 2. They cover reasoning (MAI-Thinking-1), coding (MAI-Code-1, MAI-Code-1-Flash), image generation and editing (MAI-Image-2.5, MAI-Image-2.5-Flash), speech transcription (MAI-Transcribe-1.5), and voice synthesis (MAI-Voice-2). None use distillation from OpenAI or any other external model.
Why is Microsoft building its own AI models?
Three reasons: cost, control, and competition. Microsoft pays OpenAI for every API call its products make — building its own models reduces that cost. It also reduces dependency on a single external supplier whose roadmap Microsoft doesn’t control. And OpenAI has been building its own enterprise sales products that compete directly with Microsoft Copilot, creating a conflict of interest in the relationship. Microsoft AI CEO Mustafa Suleiman described the effort as “long-term self-sufficiency for Microsoft and our partners.”
How does MAI-Thinking-1 compare to OpenAI and Anthropic models?
According to Microsoft’s own claims: MAI-Thinking-1 scores 97% on AIME 2025 and 94.5% on AIME 2026 (mathematical reasoning benchmarks), achieves 53% on SWE-Bench Pro coding benchmark — matching Claude Opus 4.6 — and was preferred over Claude Sonnet 4.6 by independent raters via Surge in blind tests. Notably, Microsoft AI CEO Mustafa Suleiman publicly acknowledged that “Anthropic still holds a lead of a few months” — an unusually candid admission at a product launch. MAI-Code-1-Flash outperforms Claude Haiku 4.5 on all four core coding benchmarks tested. These are Microsoft’s self-reported figures — independent third-party evaluations had not been published at time of writing.
Will Microsoft stop using OpenAI models?
No — Microsoft has been clear that OpenAI models will continue to be available on Azure and used in Microsoft products where they’re the best option. The MAI models are an addition to the available model portfolio, not a replacement. The relationship between Microsoft and OpenAI remains in place — but Microsoft now has genuine alternatives for standard enterprise tasks, which changes the negotiating dynamic between the two companies.
Where can developers access the MAI models?
MAI-Thinking-1 is available on Azure AI Foundry in private preview as of June 2, 2026. MAI-Code-1 is already live in GitHub Copilot and VS Code. The models will also be available on third-party platforms including Fireworks AI, Baseten, and OpenRouter for developers who don’t use Azure. Microsoft published a detailed technical report alongside the release covering training methodology and evaluation results.
What is Microsoft Scout?
Microsoft Scout is a proactive personal AI agent announced at Build 2026. Unlike Copilot — which responds to prompts — Scout runs continuously in the background across Teams and Outlook, proactively preparing meeting agendas, summarising relevant emails before you ask, and surfacing information ahead of your next task. It acts as an ambient assistant rather than a reactive chatbot.
- Microsoft Official Blog — Microsoft Build 2026: Be Yourself at Work — official MAI model announcement (June 2, 2026)
- Microsoft AI Keynote Transcript — Microsoft Build 2026 MAI Keynote — full transcript (June 2, 2026)
- CNBC — Microsoft unveils new AI models to lessen reliance on OpenAI and lower costs for developers (June 2, 2026)
- GeekWire — Microsoft unveils seven homegrown AI models in bid for ‘long term self-sufficiency’ (June 2, 2026)
- Gizbot — Microsoft Build 2026 Highlights — MAI model names and capabilities (June 2, 2026)
- Tom’s Guide — Biggest Microsoft Build 2026 announcements — live coverage (June 2, 2026)
- FelloAI — Microsoft Just Launched Its Own MAI Models — full lineup and context
- Reuters / Rappler — Microsoft teases new era of AI-driven devices at annual developer conference (June 2, 2026)
