Microsoft MAI: The 7 In-House Models From Build 2026 and the Slow OpenAI “Divorce”
On June 2, 2026, on the Build stage in San Francisco, Microsoft did what it had been rehearsing for nearly two years: it stopped depending on OpenAI for frontier models. The company unveiled the Microsoft MAI family — seven in-house models trained from scratch, spanning reasoning, code generation, image creation, voice, and transcription. These aren’t tweaks on top of someone else’s tech: per Microsoft, all of them were trained on commercially licensed data, with no distillation from any other AI lab.
The message is blunt. Mustafa Suleyman, CEO of Microsoft AI, summed it up in an interview during the event: his division was “set free” from its contract with OpenAI about six months ago to formally pursue what he calls superintelligence. For anyone who codes, uses Copilot, or runs AI workloads on Azure, this isn’t corporate gossip — it’s a shift that touches cost, speed, and who sets the roadmap of the products you use every day. Let’s break it down.
What Microsoft Announced: The MAI Family in Seven Pieces
Microsoft’s play wasn’t to beat GPT-5.5 with a single giant general-purpose model. It was to assemble a specialized family, each model optimized for one task and — the central point — for cost. The seven:
1. MAI-Thinking-1 (reasoning, the flagship)
Microsoft’s first reasoning model. It’s a mid-sized model with 35 billion active parameters (mixture-of-experts architecture) and a 256K-token context window, designed for high efficiency at low cost per token. The numbers Microsoft disclosed: roughly 97% on AIME 2025 (math) and 53% on SWE-bench Pro, the software-engineering benchmark that, per Suleyman, places the model “right alongside Opus 4.6” from Anthropic. In blind tests with human raters, Microsoft says MAI-Thinking-1 was preferred over Claude Sonnet 4.6. It’s in private preview on Microsoft Foundry.
2. MAI-Code-1-Flash (code)
The model that matters most to people who program. It’s a lightweight code model with 5 billion active parameters, tuned for GitHub and optimized for fast, cheap inference. It generates code from natural-language descriptions and is already rolling out to GitHub Copilot and VS Code through the model picker. Microsoft positions it as better price-to-performance than budget-tier rivals: under Copilot’s usage-based billing, the company says, it comes in cheaper than Claude Haiku 4.5.
3 and 4. MAI-Image-2.5 and MAI-Image-2.5 Flash (image)
Microsoft’s first in-house image models to serve both text-to-image and image-to-image (editing). The company reports top placements on the LMArena leaderboard — among the best in both generation and editing. They’re already live in PowerPoint and rolling out to OneDrive.
5. MAI-Transcribe-1.5 (transcription)
A transcription model with state-of-the-art accuracy across 43 languages, with streaming transcription on the way.
6 and 7. MAI-Voice-2 and MAI-Voice-2 Flash (voice)
The evolution of the voice line, now generating speech in 15 languages with new voice options (Microsoft says more languages are coming soon). The lineage traces back to MAI-Voice-1, Microsoft’s first fully in-house model, shipped in August 2025 — capable of generating a full minute of audio in under a second on a single GPU.
The Real Argument: Cost
The number Microsoft repeated most wasn’t a benchmark — it was a price. Suleyman claimed that, when tuned to McKinsey’s evaluation standards, the MAI models achieved “the highest win rate of any model tested at roughly ten times lower cost.” The company also cited Excel-tuned versions running up to ten times more efficiently than OpenAI’s GPT on the same kind of task.
It’s worth reading this with healthy skepticism: these are internal benchmarks, in scenarios Microsoft chose, and “preferred by human raters” isn’t the same as topping an independent public leaderboard. But the direction is clear and plausible. Smaller, specialized models running on Microsoft’s own infrastructure have a radically different cost structure than paying licensing royalties to a third-party frontier model on every API call.
And here’s the strategic kicker: running on Microsoft’s own data centers, with its own silicon, eliminates payments to outside vendors — read: OpenAI — across a growing slice of workloads. For a company processing billions of AI calls a day across Copilot, Office, GitHub, and Azure, every cent per token turns into billions on the annual bill.
The Context: The Slow Microsoft–OpenAI “Divorce”
To understand why this is big, you have to rewind. Microsoft was OpenAI’s largest investor from 2019 onward and built Copilot, Office, and much of Azure’s AI story on top of GPT. It was a comfortable dependence — until it wasn’t.
The turn started to formalize in October 2025, when the two companies restructured the partnership. OpenAI became a public benefit corporation and Microsoft kept a stake of roughly 27%, valued around $135 billion. The crucial point: Microsoft lost the right to be OpenAI’s automatic default cloud (OpenAI, in turn, committed to buy $250 billion in Azure capacity), and — per Suleyman — Microsoft’s AI division was freed to pursue superintelligence on its own, with its own researchers and infrastructure.
In November 2025, Microsoft stood up the “MAI Superintelligence Team” under Suleyman, the DeepMind co-founder who joined the company in 2024 and brought the Inflection AI team with him. In April 2026, a new chapter: OpenAI renegotiated revenue sharing, capping payments to Microsoft. The relationship didn’t end — OpenAI remains the frontier-model partner and Microsoft keeps extended IP rights — but it stopped being an exclusive marriage. It’s a slow, civil divorce, with shared custody of the products.
This power realignment echoes what we saw in OpenAI’s record fundraise, where Microsoft itself reinvested even while diversifying — read our analysis of OpenAI’s $852 billion valuation to see both sides of the table.
What This Means for Copilot and Azure
For the end user, the change is almost invisible by design. Microsoft’s strategy is to route each task to the right model: MAI-Code-1-Flash for autocomplete and simple coding; bigger models (its own or OpenAI’s) for heavy reasoning. You still see “Copilot” — under the hood, who answers shifts based on cost, speed, privacy, and customization needs.
On Azure and Foundry, developers get more options in the model picker. MAI-Code-1-Flash landing in Copilot and VS Code arrives at a sensitive moment: just days earlier, on June 1, 2026, GitHub moved Copilot to usage-based (token) billing, reigniting fears of cost spikes on agentic workflows. An ultra-efficient, cheap code model is a direct answer to that anxiety.
On Windows, Microsoft is pushing agents like Scout, integrated into Teams, Outlook, and the desktop. The cheaper inference gets, the more aggressively the company can embed AI everywhere without blowing up margins. It’s the logic of whoever controls the model factory.
What Developers Get
Stripping out the corporate noise, here’s the concrete part for people who write code:
1. Cheaper code in your day-to-day flow. A 5-billion-parameter model tuned for GitHub means autocomplete, boilerplate, and simple refactors at a fraction of the cost of calling a frontier model. For teams running Copilot at scale, the savings are real.
2. More competition, better prices. With Microsoft, OpenAI, Anthropic, and Google fighting inside the same model picker, pricing pressure favors the developer. Knowing which model to use for which task is now a skill — our comparison of the best AI model in 2026 helps you choose.
3. Specialized models as the default. Microsoft’s bet confirms a trend: the future isn’t one giant model for everything, it’s the right model for each task. Whoever can orchestrate this “fleet” of models — including the autonomous AI agents that chain several of them together — comes out ahead.
4. Less lock-in (in theory). More heavyweight vendors means more leverage to negotiate and migrate. In practice, Microsoft wants to keep you inside its ecosystem — so the freedom gain is relative.
The Risks and Caveats
Not everything is good news for those on the outside:
- Marketing benchmarks. Microsoft’s comparisons against Sonnet 4.6, Opus 4.6, and GPT are largely internal and run in favorable scenarios. “Preferred by human raters” and “tuned to McKinsey’s standards” don’t replace independent public leaderboards. Treat the numbers cautiously until third-party verification.
- Maturity. MAI-Thinking-1 is in private preview. Broad availability, final pricing, and production stability are still unknowns.
- Concentration of power. The Microsoft–OpenAI divorce doesn’t democratize AI: it just swaps one giant for another with in-house models. The AI frontier stays in the hands of a half-dozen companies. That’s the kind of scenario that draws regulatory attention — worth tracking is US AI regulation in 2026.
- Fragmentation. More models in the picker is good for price but raises complexity. Picking wrong can cost dearly in quality or in your API bill.
FAQ
What is the Microsoft MAI family?
MAI (Microsoft AI) is Microsoft’s line of in-house models, trained from scratch internally. At Build 2026, the company announced seven models covering reasoning (MAI-Thinking-1), code (MAI-Code-1-Flash), image (MAI-Image-2.5 and Flash), transcription (MAI-Transcribe-1.5), and voice (MAI-Voice-2 and Flash).
Did Microsoft break up with OpenAI?
It didn’t break up, but it sharply reduced its dependence. The partnership was restructured in October 2025 (Microsoft holding ~27% of OpenAI) and had revenue capped in April 2026. OpenAI remains the frontier-model partner, but Microsoft can now pursue superintelligence on its own and route workloads to its MAI models when cost, speed, or privacy matter more.
Is MAI-Code-1-Flash available yet?
It’s rolling out to GitHub Copilot and VS Code through the model picker. It’s a 5-billion-parameter model built for fast, cheap code. MAI-Thinking-1, the reasoning model, is in private preview on Microsoft Foundry.
Are MAI models better than GPT-5.5 or Claude?
By Microsoft’s own benchmarks, MAI-Thinking-1 ties Opus 4.6 on code (53% on SWE-bench Pro) and beats Sonnet 4.6 in blind tests. McKinsey, per Microsoft, saw the models win at roughly ten times lower cost. These are internal numbers — wait for independent verification before drawing conclusions.
Why does Microsoft want its own models?
Three reasons: cut the cost of paying OpenAI royalties on every call, gain strategic autonomy (run everything on its own data centers and silicon), and embed AI into Copilot, Office, and Windows with sustainable margins. It’s long-term self-sufficiency.
Does this make AI cheaper for developers?
It tends to, on common tasks. A 5-billion-parameter code model costs far less than a frontier model for autocomplete and simple generation. And more competition in the model picker pushes prices down across the board.
What to Watch From Here
The MAI family is a declaration of independence — but declarations have to be backed by execution. Three things to monitor in the coming months:
- Independent benchmark verification. Once MAI-Thinking-1 leaves preview and lands in public rankings (LMArena, official SWE-bench), we’ll know whether the marketing numbers hold up.
- Final pricing and availability on Foundry and Azure. The “ten times cheaper” promise only becomes a real advantage when public pricing shows up.
- The next move in the OpenAI divorce. Each renegotiation pushes the two companies closer to being direct rivals. How OpenAI responds — pricing, exclusivities, new cloud partners — defines the next chapter.
To keep up with the next updates, subscribe to the iabrief newsletter — one email per week, no hype, only what matters.
