Microsoft Breaks From OpenAI, Ships Seven MAI Models at Build 2026

For three years, Microsoft's AI products ran almost entirely on OpenAI models. At Build 2026, Mustafa Suleyman walked on stage and changed that. Microsoft launched seven new in-house AI models at Build 2026, marking the company's most significant push yet to build its own frontier AI stack independent of OpenAI. The lineup spans reasoning, coding, image generation, transcription, and voice , and every single model was built from scratch, with no shortcuts.

Zero distillation, clean slate

The phrase that keeps appearing in Microsoft's documentation is "zero distillation." Most smaller models are trained by learning to imitate a larger, more capable model , a technique called knowledge distillation. Most smaller models are trained by learning to imitate a larger, more capable model. MAI-Thinking-1 was trained without doing this. Microsoft argues that distilled models are fundamentally bound to the design choices of their teacher model and struggle to generalize to new situations. Training from scratch on their own data means the model has to genuinely learn reasoning rather than mimicking it.

Every new MAI model is trained from scratch using "clean and appropriately licensed data, without distillation from third-party models" , language that directly addresses the intellectual property concerns that have accompanied third-party model licensing. Microsoft is also co-designing models with its own Maia 200 silicon, and says it is already seeing a 1.4x efficiency boost from that work.

The flagship: MAI-Thinking-1

MAI-Thinking-1 is based on a sparse mixture-of-experts architecture with 35 billion active parameters out of approximately one trillion total parameters and a 256,000-token context window. The MoE architecture (Mixture of Experts) works by selectively activating only the subset of the model's parameters needed for each request , so you get the capacity of a trillion-parameter model at the inference cost of a 35B one.

The 256,000-token context window is enough, Microsoft said, to process a 600-page document in a single pass. On benchmarks, the numbers are strong:

• On SWE-Bench Pro, a demanding software engineering benchmark, MAI-Thinking-1 scores 53%, matching Anthropic's Claude Opus 4.6.
• On the mathematics benchmark AIME 2025, it reaches 97%, and 94.5% on AIME 2026.
• Independent human raters on Surge prefer it for overall quality in blind side-by-sides versus Sonnet 4.6.

One important caveat: Microsoft has published a preprint describing the evaluation methodology, but full reproduction of the results by independent labs has not yet occurred, leaving those benchmark claims open to challenge until confirmed externally.

The training philosophy behind the model is what Microsoft calls a "Hill-Climbing Machine" , a co-designed pipeline where every component is iteratively improvable. The reinforcement learning (RL) framework is specifically optimized for sustained log-linear improvement over many thousands of training steps, rather than the typical approach of fast early gains that plateau. Safety rewards are trained inside the same RL loop as capability rewards, so safety scales alongside performance rather than being bolted on afterward.

MAI-Thinking-1 is now available in private preview through Microsoft Foundry. It supports function calling, multi-layered instruction following, and is compatible with the widely used Chat Completions API.

MAI-Code-1-Flash: built for Copilot

MAI-Code-1-Flash is a 5-billion-parameter coding model that outperforms Claude Haiku 4.5 by 16 percentage points on SWE-Bench Pro while using 60% fewer tokens on complex tasks. That token efficiency is the real story , it means lower cost at scale for the kind of high-frequency, agentic coding tasks that Copilot handles all day.

The benchmark breakdown versus Claude Haiku 4.5:

• SWE-Bench Verified: 71.6 vs 66.6
• SWE-Bench Pro: 51.2 vs 35.2 (+16 pts)
• Terminal Bench 2: 54.8 vs 41.6

The model is rolling out to GitHub Copilot individual users in Visual Studio Code in the model picker and under the default auto picker. It was designed for agentic coding in real developer environments, and features adaptive thinking , staying concise for simple requests while spending more reasoning budget on complex tasks.

The rest of the family

The five remaining models round out a full multimodal stack:

• MAI-Image-2.5 + Flash: Now available for developers via Microsoft Foundry. The model beats Google's Nano Banana Pro on Arena's latest text-to-image leaderboard, and is already available in PowerPoint and rolling out to OneDrive. The Flash variant is priced at $1.75 per 1M tokens for text and image input, and $33 per 1M tokens for image output.
• MAI-Transcribe-1.5: Covers 43 languages with a word error rate of 2.4% (third place on the Artificial Analysis leaderboard), transcribing one hour of audio in under 15 seconds , up to 5x the speed of Gemini 3.1, Scribe v2, and GPT-4o-Transcribe. Pricing starts at $0.36 per hour.
• MAI-Voice-2: Positioned as a multilingual and more emotionally expressive successor to MAI-Voice-1, which could matter most in Copilot, Teams, Azure Speech, and voice-agent scenarios. It supports 15 languages with emotion control (excited, whispered, embarrassed), code-switching for Hindi-English and Spanish-English, and stable speaker identity across long-form content. Available through Azure Speech, with pricing starting at $22 per 1M characters.
• MAI-Voice-2-Flash: Coming soon, built for real-time agents and high-volume workloads where latency and cost matter most.

Where to use them

MAI models are also available on Fireworks AI, Baseten, and OpenRouter , not just Azure. This is a deliberate distribution move: Microsoft is making it easy for developers who don't want to go through Azure to still access the models. For the first time, developers will also be able to tune the weights of the models themselves through a feature called Microsoft Frontier Tuning.

Microsoft says a MAI model tuned for Excel already matches GPT 5.4 while being up to 10x more efficient. The claim is that custom tuning on your own workflow data produces models that are both better and cheaper than using a general-purpose frontier model.

The strategic picture

Microsoft has poured $13 billion into OpenAI; now it is building the models meant to need it less. The Build 2026 announcement is the clearest signal yet that this is a real parallel stack, not a hedge. Microsoft is building a model stack it can train, price, govern, and embed deeply into its own developer and enterprise products.

What this means for the field is a shift in what "winning" looks like for a model. For many teams, the winning model is no longer just the one with the highest public benchmark score. It is the one that fits the workflow, runs at the right cost, follows enterprise controls, and shows up where developers already work. Microsoft is betting its entire MAI strategy on that being true.

The models are available now across Microsoft Foundry, with MAI-Thinking-1 in private preview and MAI-Code-1-Flash already rolling out inside GitHub Copilot and VS Code. The MAI Playground at playground.microsoft.ai is the fastest way to try them without any infrastructure setup.