$13 billion into OpenAI. Up to $5 billion into Anthropic. Years of reselling other companies' frontier models through Azure. And then, at Build 2026 in San Francisco, Microsoft walked out and said: we built our own.
Seven in-house AI models. A new agent stack. A unified intelligence layer. A quantum chip upgrade. This wasn't a product refresh. It was Microsoft telling the market it no longer wants to only rent intelligence from the companies it funded.
Here's what was actually announced, what the numbers say, and where the claims need more scrutiny.
- MAI Thinking One: The Reasoning Model Microsoft Built From Scratch
- The Full MAI Family: All 7 Models and What They Cover
- Microsoft IQ: The Intelligence Layer Behind the Agents
- Scout: Microsoft's First Always-On Autonomous Agent
- M-dash: 100+ Agents Hunting for Security Bugs in Your Code
- Majorana 2: The Quantum Chip With a 2029 Promise
- The Real Story: Investor, Partner, and Now Competitor
- FAQ
MAI Thinking One: The Reasoning Model Microsoft Built From Scratch
MAI Thinking One is Microsoft's first in-house reasoning model and the centerpiece of the Build 2026 launch. It has 35 billion active parameters and was trained on what Microsoft describes as clean, commercially licensed data, with no distillation from third-party model outputs.
That training claim deserves attention. Several AI companies have faced questions about whether their models were trained on outputs from competitors' systems, which raises both legal and business risks. Microsoft is drawing a clean line here, positioning MAI Thinking One as a model it owns fully, without a dependency chain running back to OpenAI or anyone else.
On performance, Microsoft AI CEO Mustafa Suleyman said that after tuning MAI models for McKinsey, Microsoft outperformed GPT-4.5 on quality while projecting roughly 10x better cost efficiency based on public pricing data scaled across model sizes. The word "projecting" matters here. This is not a measured production result. It is an estimate derived from comparing public pricing at different model scales.
On coding specifically, Microsoft says MAI Thinking One was preferred over Claude Sonnet 4.6 in blind evaluations run by Serge, described as an independent human rating partner. The company also says the model matches Claude Opus 4.6 on SWEBench Pro, a widely used software engineering benchmark. Neither the Serge evaluation methodology nor the exact SWEBench Pro scores were disclosed in the announcement.
The model is currently in private preview on Microsoft AI Foundry.
The Full MAI Family: All 7 Models and What They Cover
Microsoft's announcement covers a broad surface area. These are not seven variations of the same model. They span different modalities and workflows:
MAI Thinking One is the flagship reasoning model described above. It also has a Flash variant built for speed and efficiency at lower cost, following the now-standard pattern of pairing a heavy reasoning model with a faster, cheaper sibling for tasks that don't need maximum capability.
MAI Code One is designed to convert plain-language descriptions into source code for apps and websites. It is already rolling out inside GitHub Copilot and Visual Studio Code, meaning it is not a lab release. Millions of developers will encounter it through tools they already use daily.
MAI Image 2.5 supports both text-to-image and image-to-image generation. Users can describe an output in plain language or pass in sketches and visual references to guide the result. It is already live inside PowerPoint and is rolling out on OneDrive. A Flash variant is included here as well.
MAI Transcribe 1.5 handles high-accuracy transcription across 43 languages, with streaming support coming soon.
MAI Voice 2 and its Flash variant add support for more than 15 additional languages and new voice options for text-to-speech output.
Taken together, Microsoft is building a model layer that covers reasoning, coding, image generation, transcription, and voice. Every major surface of modern knowledge work has at least one model pointed at it.
Microsoft IQ: The Intelligence Layer Behind the Agents
Microsoft IQ is now generally available. The goal is to move AI agents away from generic chatbot behavior by grounding them in actual organizational context. Less hallucination, more usefulness inside a specific company's environment.
It is built from four components. Work IQ captures how people work inside Microsoft 365, understanding emails, documents, meetings, contacts, and the relationships between them. Its APIs become available on June 16th. Fabric IQ runs on Microsoft Fabric and acts as a semantic foundation for structured business data, giving that data a more organized meaning layer. Foundry IQ handles unstructured information like wikis, policies, contracts, and live web content. Web IQ is the newest addition, providing real-time web grounding through search. It is model-agnostic and built natively on the Model Context Protocol (MCP), which is becoming a standard way for agents to connect to external tools and data sources. Microsoft says Web IQ returns relevant information blocks nearly 2.5 times faster than the next best alternative.
The architecture matters because it is what makes Scout possible.
Scout: Microsoft's First Always-On Autonomous Agent
Scout is Microsoft's first autopilot agent, a new category the company defines as always-on agents that stay active in the background, work autonomously, and act on your behalf under the permissions you or your organization set.
The identity piece is significant. Scout does not run as an anonymous shared service account. It operates under its own governed Microsoft Entra identity, meaning every action it takes is traceable back to a known actor inside the company directory. Its credentials are scoped to the task, protected end to end, redacted from logs and diagnostics, and managed like a first-party Microsoft service.
In practice, Scout handles coordination work. It proactively schedules meetings across time zones, flags important upcoming events, generates prep materials, identifies deliverables, blocks calendar time automatically, and flags stalled decisions before they become bigger problems. It connects to Teams, Outlook, OneDrive, SharePoint, browser resources, local desktop resources, and MCP servers.
Scout is built on OpenClaw, an open-source framework that appeared in November 2025, designed to give agents a continuous operating rhythm. Microsoft is contributing policy conformance work upstream so organizations can validate whether their environment meets security and compliance requirements.
Enterprise guardrails are built in. Scout can only access approved resources. Sensitive actions can require human sign-off. Microsoft Purview policies including sensitivity labels and data loss prevention rules are enforced before anything is sent or written. Currently in private preview for Frontier organizations, access requires Frontier enrollment, Intune policy configuration, and a GitHub Copilot license.
M-dash: 100+ Agents Hunting for Security Bugs in Your Code
The name is a deliberate joke. AI-generated text is famously full of em dashes. Microsoft named its AI security system M-dash. The product underneath is serious.
M-dash is a multi-model agentic security system that deploys more than 100 specialized agents to search for exploitable bugs in code. These agents reason about data flows, business logic, exploit chains, and context-aware fixes inside the developer portal. The idea is that a swarm of specialized agents reviewing code from different angles can find problems that a standard static scanner would miss. No public release date was announced.
Majorana 2: The Quantum Chip With a 2029 Promise
Majorana 2 is Microsoft's next-generation topological quantum chip. The company claims it is 1,000 times more reliable than its previous generation of qubits. Mean qubit lifetime is now 20 seconds, with some instances lasting up to a minute. Most competing approaches measure qubit lifetime in microseconds or milliseconds. Microsoft compared the improvement to a phone battery going from dying in a day to lasting nearly three years on a single charge.
A hardware change sits behind this improvement. Majorana 1 used aluminum as a superconductor. Majorana 2 uses lead, which helps shield fragile qubits from cosmic disturbances that cause instability.
The current chip has 12 qubits. A useful quantum computer would require millions. Microsoft says the combination of reliability, one-microsecond operations, and very small qubit size (around 1/100th of a millimeter) puts it on a path toward a commercially viable, scalable quantum computer by 2029.
Microsoft Discovery, now generally available, is the AI platform the quantum team is already using. It lets teams of AI agents search, research, reason, generate hypotheses, and optimize experiments while human experts stay in control. One agent found an uncalibrated temperature sensor hidden in fabrication data. Another helped cut the cycle time for adjusting topological quantum state parameters, which previously required weeks of manual measurement, by what Microsoft describes as orders of magnitude.
The Real Story: Investor, Partner, and Now Competitor
Microsoft committed $13 billion to OpenAI and up to $5 billion to Anthropic. It sells both companies' models through Azure. It is their investor, their partner, their infrastructure provider, and now their direct competitor.
That position is unusual. And it explains why Build 2026 felt like more than a developer conference.
Every time Microsoft ran a third-party model on Azure, part of the economics went to an outside provider. With its own models running on its own infrastructure, Microsoft controls the full stack: lower costs, better margins, and the ability to offer cheaper AI tools to developers without depending on pricing decisions made elsewhere. Satya Nadella put it plainly at the conference: the time has come for every company to move from just consuming a frontier model to fully participating at the frontier.
Both OpenAI and Anthropic are reportedly moving toward IPOs. Microsoft is tied to both of them financially. The strategic tension that creates is not theoretical. It is already visible in the benchmark claims Microsoft made at this event, specifically naming Claude Sonnet 4.6 and Claude Opus 4.6 as the models MAI Thinking One was measured against.
Whether the models hold up against that framing in independent testing is a separate question. But the direction is clear. Microsoft is not building AI tools as a distribution channel for other companies anymore. It is building them to own the stack.
My Take
The benchmark claims are the part I'd hold loosely. "Preferred over Claude Sonnet 4.6 in blind evaluations" from a partner called Serge, with no published methodology and no raw scores, is not the same as a third-party independent benchmark. The 10x cost efficiency number is a projection from public pricing math, not a production result. These details matter because Microsoft's entire positioning at Build 2026 rests on the idea that its models can now compete at the frontier, and that claim needs harder evidence than what was shown on stage.
The parts that are harder to dismiss: MAI Code One shipping directly inside GitHub Copilot and VS Code is not a lab announcement. That is a real distribution move. Microsoft IQ's Work IQ APIs going live on June 16th is a real date. Scout running under governed Entra identities is a real enterprise capability, not a demo.
The Majorana 2 claims are in a different category entirely. 1,000x reliability is extraordinary, the retracted 2018 paper is a real data point, and no peer review exists yet. The 2029 target for a commercially viable quantum computer is either the most important thing Microsoft announced or the most aggressive forward guidance in the whole event. Probably worth revisiting in 12 months.
- Microsoft launched 7 in-house MAI models at Build 2026, covering reasoning, coding, image generation, transcription, and voice.
- MAI Thinking One has 35B active parameters and was trained without third-party model distillation. Its context window is listed as either 256K or 128K tokens depending on the source, and this discrepancy has not been officially clarified.
- Microsoft claims MAI Thinking One outperformed GPT-4.5 on quality for McKinsey and projects roughly 10x cost efficiency. These are projections, not measured production results.
- Scout is Microsoft's first always-on autonomous agent, currently in private preview for Frontier organizations with a GitHub Copilot license.
- Microsoft IQ is generally available. Work IQ APIs open on June 16th.
- Majorana 2 claims 1,000x qubit reliability improvement, but supporting research is not yet peer-reviewed. The 2029 commercial quantum computer target should be treated as a long-horizon claim.
- Microsoft is now simultaneously investor, partner, infrastructure provider, and direct competitor to OpenAI and Anthropic.
FAQ
What is MAI Thinking One?
MAI Thinking One is Microsoft's first in-house reasoning model, built from scratch with 35 billion active parameters. Microsoft says it was trained on commercially licensed data without distillation from third-party models. It is currently available in private preview on Microsoft AI Foundry.
How does MAI Thinking One compare to GPT-4.5 and Claude?
Microsoft claims MAI Thinking One outperformed GPT-4.5 on quality in McKinsey-tuned evaluations while projecting about 10x better cost efficiency based on public pricing comparisons. On coding, the company says it was preferred over Claude Sonnet 4.6 in blind evaluations by an independent rating partner called Serge, and that it matches Claude Opus 4.6 on SWEBench Pro. Full methodology and raw scores have not been published.
What is Microsoft Scout and how do I get access?
Scout is Microsoft's first always-on autonomous agent. It handles coordination tasks like meeting scheduling, prep materials, and calendar blocking inside Microsoft 365 apps. Access currently requires Frontier organization enrollment, Intune policy configuration, and a GitHub Copilot license. It is in private preview as of Build 2026.
Is Majorana 2 a real breakthrough?
Microsoft's claims are significant on paper. A mean qubit lifetime of 20 seconds, up from typical microsecond or millisecond lifetimes elsewhere, would represent a major step forward. However, the supporting research has not been peer-reviewed. Microsoft also had to retract a related Nature paper from 2018. The claims are worth watching but not yet independently validated.
What is the Model Context Protocol (MCP) and why does Microsoft keep mentioning it?
MCP is becoming a standard protocol for connecting AI agents to external tools and data sources. Microsoft's Web IQ and Scout both use it natively. By building MCP support into its agent stack, Microsoft is positioning its tools to work within an emerging open standard rather than a proprietary connector system.
Where This Goes Next
Build 2026 announced a direction, not a completed product. Most of what Microsoft showed is in private preview or rolling out gradually. The benchmark claims need independent confirmation. The Majorana 2 physics need peer review. The 10x cost efficiency number needs real-world production data behind it.
But the structural position is already locked in. Microsoft owns infrastructure that OpenAI and Anthropic depend on while building models that compete with them directly. That is an unusual place to be, and it is not a position any other company in the current AI market occupies quite the same way.
The more interesting question is not whether these models are better than the competition today. It is whether Microsoft can close the gap fast enough to matter before OpenAI and Anthropic reach the scale and independence that makes the investment relationship irrelevant.
0 Comments