At its annual Build developer conference on June 2 in San Francisco, Microsoft dropped one of its biggest-ever AI product blitzes — seven new in-house MAI models, a wearable AI gadget concept, new on-device models for Windows PCs, and a security-first agent containment framework. The event came just one day after Anthropic filed confidentially for an IPO, sharpening the competitive stakes across the entire AI stack. Microsoft has invested $13 billion in OpenAI and $5 billion in Anthropic — and is now building its own models to compete with both.
The shift is economic as much as technical. By running its own MAI models on Azure infrastructure, Microsoft avoids the profit-sharing costs tied to third-party model providers. For developers, that means lower per-token costs and deeper integration across consumer and enterprise surfaces alike.
Microsoft’s MAI Model Family — Seven Models, One Strategy
From reasoning and coding to images, transcription, and voice — tap each model to explore what it does, its specs, and where it ships.
The Full MAI Lineup — Tap to Explore
Each model targets a specific task. Select a model below to see what it does, where it ships, and how it stacks up.
MAI-Thinking-1 — Reasoning Model
MAI-Thinking-1 is Microsoft’s first fully homegrown reasoning model, built for high efficiency and performance at a low token cost. It is available in private preview through Microsoft Foundry, where developers and enterprises can request early access before broader rollout. Customers can improve its accuracy further by feeding in their own data through Microsoft’s Frontier Tuning process.
On the SWE Bench Pro coding benchmark — one of the toughest evaluations in software engineering — Mustafa Suleyman said the model sits “right alongside Claude Opus 4.6” in side-by-side evaluations. When tuned for McKinsey’s enterprise workflows, the model achieved the highest win rate of any tested model, outperforming GPT-5.5 at roughly 10× lower cost. A separate Excel-tuned MAI model matched GPT-5.4 while being up to 10× more efficient.
MAI-Code-1-Flash — Coding Model
MAI-Code-1-Flash is Microsoft’s inference-efficient coding model — a 5-billion parameter model trained on production GitHub Copilot workflows and licensed code data. It converts written descriptions into source code for applications and websites, a workflow now widely known as vibe coding.
The model is live today inside GitHub Copilot and Visual Studio Code. Kyle Daigle, Microsoft’s developer marketing chief and GitHub operating chief, described it as “inference ultra-efficient” in a blog post. It is trained with adaptive solution-length control, so it avoids generating bloated output for simple tasks. Microsoft compares its performance to Claude Haiku-class models at lower cost.
MAI-Image-2.5 — Image Generation & Editing
MAI-Image-2.5 and its Flash variant handle text-to-image generation and image editing. The model is already live in PowerPoint and coming to OneDrive soon. It is also available through Microsoft Foundry. Microsoft claims it outperforms Google’s Nano Banana Pro model on the ELO image-editing benchmark.
MAI-Transcribe-1.5 — Speech-to-Text
MAI-Transcribe-1.5 is a transcription model covering 43 languages. Microsoft describes it as state-of-the-art accuracy and five times faster than competing models from major hyperscalers. It is intended for developers building captioning, dictation, and audio-to-text workflows, and runs through Azure.
MAI-Voice-2 — Speech Generation
MAI-Voice-2 and its Flash variant (coming soon) cover natural speech generation in 15 additional languages. The model can adapt to a speaker’s voice from a short audio sample — a feature Microsoft says includes protections specifically against voice cloning abuse, built in from the start. It powers synthetic voice output across Copilot and developer surfaces.
Aion 1.0 — On-Device Models for Windows
Aion 1.0 is Microsoft’s new generation of small language models built for local execution on Windows. There are two variants. Aion 1.0 Instruct handles everyday text tasks — summarisation, rewrites, intents, and accessibility — without a cloud round-trip. It previews in Microsoft Edge Insider channels now and arrives as an open-source model on Hugging Face in July.
Aion 1.0 Plan is a 14-billion parameter reasoning and tool-calling model with a 32K context window. It ships in-box on capable Windows devices and enables fully local agentic workflows — managing files, invoking tools, and orchestrating sub-agents — entirely offline with no per-token cost.
Why Microsoft Is Building Its Own Models Now
For years, Microsoft’s AI strategy relied heavily on OpenAI — providing cloud infrastructure and reselling access through Azure and Copilot. That dynamic has been shifting. In April 2026, Microsoft and OpenAI amended their partnership agreement, ending Microsoft’s exclusive access to OpenAI’s intellectual property and allowing OpenAI to sell models on competing cloud platforms including Amazon Web Services.
The economic logic behind the MAI family is straightforward. Every call to an OpenAI or Anthropic model via Azure carries a profit-sharing cost. By substituting proprietary MAI models for routine workloads, Microsoft avoids those costs entirely while keeping customers on Azure. As Mustafa Suleyman put it in the primary Microsoft AI announcement: “Developers and businesses have been crying out for AI that delivers on their terms and under their say.”
Meanwhile, Anthropic filed its confidential S-1 with the SEC on June 1, one day before Build, and OpenAI is also preparing a public offering. With both major model partners moving toward public markets and greater commercial independence, Microsoft’s decision to build its own model stack carries long-term supply-chain logic beyond the immediate cost savings. The company has backed wearable tech ambitions before — its history with HoloLens is a relevant reference point for how hardware concepts can evolve.
The Road to Microsoft’s AI Independence
Key moments leading up to the Build 2026 announcements.
The Agent-First Hardware Concept
Steven Bathiche, CVP and Technical Fellow of Microsoft’s Applied Sciences Group, previewed two concept devices designed for AI agent interactions in workplace settings. These are reference designs, not confirmed commercial products.
Smart Badge (Wearable)
A small wearable device the size of an ID card, designed to hang around the neck or on a belt loop. Activated by fingerprint, it gives workers quick access to AI agents outside of a laptop or desktop. It includes a camera so agents can “understand and help take action on the environment around them,” as Bathiche wrote in his blog post.
- Qualcomm wearable chip
- 5G connectivity
- Touch and voice interface
- Fingerprint activation
- Camera for agent vision
- Runs on Android (MDEP platform)
Desk Hub (Stationary)
A small desktop cube with a touch and voice-activated screen. It functions as a companion to an existing PC and can connect to cloud PCs via Windows 365. Satya Nadella described the concept as representing “a new form factor” for computing.
- MediaTek IoT silicon
- Touch and voice interface
- Windows 365 integration
- Connects to existing PC
- Runs on Android (MDEP platform)
- Enterprise manageability built in
What you just saw is a pretty significant shift. We believe the time has come for every company to move from consuming a frontier model to fully participating at the frontier in the frontier ecosystem.
Developers and businesses have been crying out for AI that delivers on their terms and under their say. We see this as a major step towards delivering that.
Continuously-running local agents, like Hermes Agent, require intentional isolation. Developers need control over what an agent can access and trust that those controls will hold.
What Changed for Developers on Windows
Alongside the MAI model family, Microsoft announced a slate of developer-focused updates to Windows 11. Coreutils for Windows — a cross-platform reimplementation of GNU Coreutils in Rust — is now generally available, giving developers Linux-compatible command-line utilities that run natively on Windows. WSL containers, bringing built-in Linux container support to Windows, are heading to public preview.
On the security side, Microsoft Execution Containers (MXC) is now in early preview — a policy-driven execution layer that lets developers declare what an AI agent can access (files, network, etc.) and enforces those boundaries at runtime. Partners including OpenAI, OpenClaw, Manus, Hermes, and NVIDIA are already integrating MXC into their agent frameworks.
On the hardware front, the Surface RTX Spark Dev Box — powered by NVIDIA RTX Spark silicon — delivers 1 petaflop of AI compute (FP4, with sparsity) and 128 GB of unified memory shared across CPU and GPU. It ships later this year, exclusively from Microsoft.com in the US. The DGX Station for Windows, built on NVIDIA’s GB300 Grace Blackwell Ultra Superchip, arrives in Q4 2026 and can run frontier AI models up to 1 trillion parameters locally — making it the most powerful deskside developer workstation ever available for Windows. For those tracking the NVIDIA RTX Spark ecosystem, this is where silicon and software strategy converge.
Gaming and security enthusiasts can note that post-quantum cryptography (PQC) support is expanding across the Windows platform, including PQ hybrid key exchange in the Windows TLS stack and PQ certificate issuance via Active Directory. Driver signing also moves toward a higher security bar under the updated Windows Hardware Compatibility Program. Developers interested in how security investments affect hardware performance can follow the ongoing Windows 11 platform updates.
Microsoft Build 2026 covered the launch of seven new in-house MAI models — spanning reasoning, coding, image generation, transcription, and voice — along with on-device Aion SLMs for Windows, Project Solara concept hardware, Microsoft Execution Containers for agent security, and a range of developer platform updates including Coreutils, WSL containers, and Windows 365 for Agents. Details on all announcements are available via the official Windows Developer Blog and the Microsoft AI model launch post.
The Surface RTX Spark Dev Box and DGX Station for Windows were also discussed as part of the developer hardware rollout, with the former available later this year and the latter in Q4 2026. The wearable badge and desk hub concepts under Project Solara were described as reference designs currently in internal pilot with a few hundred Microsoft employees.
Related coverage: Microsoft 365 security threats and MFA bypass risks · Wearable tech in 2026
