Microsoft’s Build 2026 conference delivered a move that had been anticipated for months but still landed with weight: the company is no longer content to sit behind its $13 billion investment in OpenAI and resell frontier models through Azure. The Microsoft AI Superintelligence team announced seven new in-house models under the MAI (Microsoft AI) banner, including the company’s first reasoning model and its own coding model now powering GitHub Copilot. This isn’t an incremental update — it’s a strategic pivot that changes the calculus for every developer building on Azure.
The lineup spans reasoning, coding, image generation, transcription, and voice synthesis. Here’s what each model brings to the table and why it matters beyond the headlines.
MAI-Thinking-1: Microsoft’s First Reasoning Model
The flagship of the announcement is MAI-Thinking-1, a medium-sized reasoning model built on a sparse Mixture of Experts architecture with 35 billion active parameters out of roughly 1 trillion total. It offers a 256K token context window and supports function calling through a Chat Completions-compatible API.
What makes MAI-Thinking-1 noteworthy isn’t just the architecture — it’s the training philosophy. Microsoft trained this model entirely from scratch on clean, commercially licensed data, explicitly excluding AI-generated content from the pre-training corpus. There was no distillation from third-party models. The team argues this approach produces models that are more steerable and adaptable than those built by imitating existing systems, since an “imitator is fundamentally tied to the design choices of its teacher.”
On benchmarks, MAI-Thinking-1 lands in striking distance of the best models in its weight class. It scores 97.0% on AIME 2025 and 94.5% on AIME 2026 for mathematical reasoning. On SWE-Bench Pro, Microsoft reports that it matches Anthropic’s Claude Opus 4.6 on coding tasks — a strong result for a model significantly smaller than the dense alternatives. In blind human evaluations across 1,276 tasks covering single-turn and multi-turn conversations, professional raters preferred MAI-Thinking-1 over Claude Sonnet 4.6.
The model is currently available in private preview through Microsoft Foundry, with a public preview on the MAI Playground coming soon.
MAI-Code-1-Flash: The Coding Model Behind GitHub Copilot
MAI-Code-1-Flash is a 5-billion active parameter coding model tuned specifically for agentic coding workflows. It’s already live in GitHub Copilot and Visual Studio Code — marking the first time Microsoft has used its own model rather than OpenAI’s to power the core coding experience in Copilot.
The model is designed for inference efficiency, which matters when you consider that Copilot serves millions of developers making thousands of completions per session. At 5B active parameters, it sits in the “small but capable” category, achieving 51% on SWE-Bench while keeping inference costs low enough for high-volume usage.
What sets MAI-Code-1-Flash apart from generic coding models is its deep integration with the Microsoft developer stack. It’s been trained on GitHub-specific patterns, VS Code extension APIs, and the typical workflows developers follow in the Microsoft ecosystem. This isn’t just a model that writes code — it’s a model that understands the context in which that code gets written.
The Broader MAI Family
Beyond reasoning and coding, the Build announcements included updates across the multimodal stack:
MAI-Image-2.5 (with a Flash variant) handles text-to-image generation and editing. It’s already integrated into PowerPoint and rolling out to OneDrive. Microsoft claims it surpasses Google’s Nano Banana 2 on the Arena image quality leaderboard — a credible claim given the model is production-deployed, not just a paper artifact.
MAI-Transcribe-1.5 brings speech-to-text accuracy across 43 languages, up from 25 in the previous version. It runs 5× faster than competing models and includes built-in support for domain-specific terminology — useful for medical, legal, and technical transcription where accuracy on jargon matters more than general language fluency.
MAI-Voice-2 adds natural speech generation across 15 languages with new voice options, including the ability to adapt to a voice from a short sample. A Flash variant is coming for lower-cost deployments.
Frontier Tuning: Enterprises Train Their Own Models
Perhaps the most architecturally interesting announcement isn’t a model at all — it’s Frontier Tuning, a paradigm where enterprises use reinforcement learning in their own environments to fine-tune MAI models on real workflows. The concept works through “Reinforcement Learning Environments” (RLEs) — private training loops where models learn from the actual sequence of steps, decisions, and actions that happen in a company’s daily operations.
Microsoft reports that a MAI model tuned for Excel workflows matches OpenAI’s GPT-5.4 performance while running at roughly 10× lower cost. When tuned for a large enterprise’s internal standards, the model achieved the highest win rate of any model tested, again at roughly 10× lower cost. The key insight: domain-specific tuning on proprietary data can close — or even reverse — the gap between a mid-sized first-party model and the largest third-party frontier models.
For developers, this means the model you deploy might perform very differently from the base model benchmarks suggest. The tuning story matters as much as the base model capabilities.
Using MAI Models: Chat Completions and Function Calling
All MAI models expose a Chat Completions-compatible API with function calling support. Since the models are available through OpenRouter, Fireworks AI, and Baseten in addition to Microsoft Foundry, integrating them into existing applications requires minimal code changes. Here’s what a function-calling integration looks like with an OpenAI-compatible endpoint:
import openai
client = openai.OpenAI(
base_url="https://openrouter.ai/api/v1",
api_key="YOUR_OPENROUTER_API_KEY",
)
tools = [
{
"type": "function",
"function": {
"name": "lookup_customer",
"description": "Fetch customer details by ID",
"parameters": {
"type": "object",
"properties": {
"customer_id": {
"type": "string",
"description": "The customer identifier",
}
},
"required": ["customer_id"],
},
},
}
]
response = client.chat.completions.create(
model="microsoft/mai-thinking-1",
messages=[
{"role": "system", "content": "You are a customer support assistant."},
{"role": "user", "content": "Look up customer acme-42 and summarize their recent orders."},
],
tools=tools,
tool_choice="auto",
)
for tool_call in response.choices[0].message.tool_calls:
print(f"Calling: {tool_call.function.name}")
print(f"Arguments: {tool_call.function.arguments}")
The function calling support means MAI-Thinking-1 can participate in agentic workflows — routing requests to the right tools, composing multi-step actions, and recovering from errors. That’s the pattern where reasoning models deliver the most value in production systems.
What This Means for the AI Landscape
Microsoft’s move into first-party frontier models creates a new dynamic. Azure customers now have a genuine alternative to OpenAI and Anthropic models that’s optimized for Microsoft’s own infrastructure. The co-design with Maia 200 custom silicon already delivers a reported 1.4× efficiency boost, and the “no distillation” training philosophy means these models aren’t derivative — they represent a separate lineage with its own strengths and failure modes.
For developers, the practical takeaway is straightforward. If you’re building on Azure or using GitHub Copilot, MAI models will increasingly be the default — and with Frontier Tuning, you can shape them to your specific domain without waiting for the next base model release. The models are available now through Microsoft Foundry and third-party platforms, and MAI-Code-1-Flash is already shipping in VS Code.
The competitive pressure on pricing and capabilities will benefit everyone. When a company that hosts both OpenAI and Anthropic models starts offering its own competitive alternatives, the economics of frontier AI shift — and developers are the ones who benefit most.