Microsoft Bets on Seven Homegrown AI Models to Loosen OpenAI Grip

Microsoft unveiled seven of its own AI models at its Build developer conference in San Francisco on June 2, led by MAI-Thinking-1, its first reasoning system. The company built them in-house to make its dependence on OpenAI optional, a strategic hedge that arrives even as it keeps a licence to OpenAI’s technology that runs into the next decade.

It’s a wide swing across models, agents and hardware. The flagship reasoning model also lands roughly a year and a half after OpenAI and Google shipped theirs, which is the catch in an announcement Microsoft wants read as proof it can stand on its own.

Microsoft’s Homegrown Lineup Arrives in Foundry

The seven-model family came from Microsoft AI, the consumer and research unit run by Mustafa Suleyman, chief executive of Microsoft AI and a co-founder of DeepMind. He described the launch as a push for long-term self-sufficiency in Microsoft AI’s announcement of the seven new models, the company’s clearest statement yet that it wants its own frontier-adjacent technology, not just OpenAI’s.

Here is what shipped:

• MAI-Thinking-1: the first reasoning model, with about 35 billion active parameters, aimed at math, science and coding problems
• MAI-Code-1-Flash: a 5-billion-parameter coding model already rolling out inside Visual Studio Code and GitHub Copilot
• MAI-Image-2.5 and MAI-Image-2.5 Flash: two image-generation models
• MAI-Transcribe-1.5: an audio transcription model
• MAI-Voice-2 and MAI-Voice-2 Flash: synthetic speech models

Microsoft is selling them on cost as much as capability. The 5-billion-parameter coder is pitched as comparable to a small rival model at a lower price, and the whole family is reaching developers through Azure AI Foundry, the company’s model-hosting service. The point is to give Microsoft’s own customers a cheaper, in-house option sitting right next to the OpenAI models they already use.

Why the OpenAI Leash Still Runs to 2032

Microsoft was the first big company to pour money into OpenAI, and untangling that has taken most of a year. The two sides reworked their alliance in 2025, and the new shape is a partnership Microsoft no longer fully controls.

Microsoft now holds a non-exclusive licence to OpenAI’s technology that runs to 2032, keeps a 27% equity stake after converting its investment at a $135 billion valuation, and stays a primary cloud provider, with OpenAI committed to spending $250 billion on Azure. A separate April change capped the revenue OpenAI shares with Microsoft at $38 billion, terms set out in the company’s regulatory filing on the restructured OpenAI partnership. OpenAI, meanwhile, won the freedom to sell across rival clouds, including Amazon and Google.

So the dependence is real and it is contractual. That is the backdrop for the in-house push, and a senior Microsoft AI hire put the reasoning bluntly.

It’s important that we are self-sustaining and are not taking huge dependencies, as this is a very fast-moving, highly fluctuating environment.

Sophie Lebrecht, who joined Microsoft’s AI team in March, said that during a press visit to the company’s Silicon Valley campus.

A Reasoning Model That Lands a Year and a Half Late

The headline product is also the most exposed. MAI-Thinking-1 is a reasoning model, the kind that works through a problem step by step before answering, and it arrives roughly a year and a half behind the systems OpenAI and Google pioneered. For now it is limited to a select group of customers.

Microsoft says it built the model “from scratch” with “no distillation,” the shortcut where a developer trains a cheaper system by copying a stronger rival’s outputs. The clean-data pitch is aimed squarely at enterprise buyers worried about where a model’s training came from. Outside reviewers who read the technical notes within a day argued the corpus is mostly web-crawled, much like other large models, which takes some shine off the licensed-data framing.

On its own scorecard, the numbers are competitive:

• 97.0% on AIME 2025, the American Invitational Mathematics Examination used to test multi-step math reasoning
• 94.5% on the 2026 edition of the same exam
• A tie with Claude Opus 4.6 on SWE-Bench Pro, a software-engineering benchmark from Anthropic’s model line
• A win over Claude Sonnet 4.6 with human raters in Microsoft’s own blind side-by-side tests

Read those carefully. They are Microsoft’s figures, and the comparisons are against a much larger rival model with a fraction of the active parameters. Impressive for the weight class, short of the absolute frontier.

Scout Turns a Viral Agent Into Office Software

The more surprising launch was Microsoft Scout, an always-on assistant that prepares meetings, manages schedules and drafts email. It is built on OpenClaw, the open-source agent software whose popularity kicked off the current agent wave late last year. Scout connects to Teams, Outlook, OneDrive and SharePoint, then works in the background, scheduling across time zones, flagging stalled decisions and blocking calendar time before deadlines. A built-in conformance system checks its own behaviour and leaves an audit trail.

The optics were striking. OpenClaw came from Austrian developer Peter Steinberger, gathered 180,000 GitHub stars in about three months, and got its creator hired by OpenAI in February. Nadella once called the autonomous-agent phenomenon “a virus,” rattled by early security incidents. On Tuesday, Steinberger walked onto the Build stage to applause from the same executives.

Microsoft is not alone in chasing this. Here is how the three agents line up:

Google unveiled Gemini Spark, its own autonomous agent, last month for paying US users. Scout is reserved for a limited circle for now too. The race is on, and Microsoft’s entry runs on code an outsider wrote.

An Offline Mini-PC and a Home Full of Agents

The bet extends into hardware. Microsoft showed the Surface RTX Spark Dev Box, an Nvidia-powered mini-PC for developers who want to run big models without paying for the cloud.

• Surface RTX Spark Dev Box: built on Nvidia’s RTX Spark superchip, delivering up to 1 petaflop of AI compute with 128GB of unified memory, enough to run models above 120 billion parameters locally; US only, later this year, with industry estimates near $3,000 to $3,500
• A desk speaker with a screen: it recognises you by face, shows your day’s tasks, and doubles as a computer once plugged into a monitor
• A wearable badge: built with Qualcomm, for talking to your AI agent by voice without opening apps

The mini-PC can load and run those models without a single cloud API call, according to Microsoft’s Surface RTX Spark Dev Box hardware briefing. The two home prototypes point at a different idea: a line of Android-based devices you talk to, where the AI agent acts for you instead of you tapping through software. Microsoft also announced an AI platform aimed at scientific research.

Nadella’s IBM Fear Is Driving All of It

Satya Nadella has repeated the same warning for years. He refuses to let Microsoft become the next IBM, the computing giant that bankrolled his company’s rise in the 1980s and then lost the lead to it. The lesson he draws is about dependence: lean too hard on one supplier of the core technology, and you hand that supplier your future.

The seven models, the agent, the Nvidia box and the home devices all point the same way. Microsoft wants to own enough of the stack that no single partner can dictate terms, and Lebrecht’s campus comment about self-sufficiency was the logic stated out loud.

Owning the stack and leading it are different things. MAI-Thinking-1 trails the frontier, the clean-data claim drew scrutiny within a day, and the most useful agent on stage runs on software an outsider wrote. Microsoft still holds its OpenAI licence and its stake in the company; the seven models are its argument that, when it matters most, it can stand without leaning on either.

Frequently Asked Questions

What Is Microsoft’s MAI-Thinking-1?

MAI-Thinking-1 is Microsoft’s first reasoning model, a system that works through problems step by step before answering. It runs about 35 billion active parameters and targets math, science and coding tasks, and Microsoft says it was trained without distilling rival models’ outputs.

Can Developers Use the New MAI Models Now?

Partly. MAI-Code-1-Flash, a 5-billion-parameter coding model, is rolling out inside Visual Studio Code and GitHub Copilot, and the image, voice and transcription models are reaching Azure AI Foundry. MAI-Thinking-1 is limited to a select group of customers for now.

Is Microsoft Ending Its Partnership With OpenAI?

No. Microsoft keeps a non-exclusive licence to OpenAI’s technology, a 27% equity stake and its role as a primary cloud provider. The new models give customers a second route to advanced AI inside Microsoft’s stack rather than replacing OpenAI.

How Much Will the Surface RTX Spark Dev Box Cost?

Microsoft has not set an official price. Industry estimates put it between $3,000 and $3,500. The Nvidia-powered mini-PC, detailed on Microsoft’s Surface RTX Spark Dev Box product page, will sell in the United States later this year on the company’s own store.

What Is OpenClaw and How Does Microsoft Scout Use It?

OpenClaw is an open-source personal AI agent released in January and built by Austrian developer Peter Steinberger; it gathered about 180,000 GitHub stars in three months. Microsoft Scout is an office assistant built on the same framework, wired into Teams, Outlook, OneDrive and SharePoint.