TL;DR Microsoft MAI Models are Microsoft's first fully in-house AI model family, covering reasoning, coding, image generation, transcription, and voice. The standout releases are MAI-Thinking-1, a 1T-parameter MoE reasoning model in private preview, and MAI-Code-1-Flash, a 5B coding model available now in GitHub Copilot. Beyond the benchmarks, the launch signals Microsoft's growing independence from OpenAI and gives enterprises a new Microsoft-native AI option.

Microsoft spent years as OpenAI's largest backer and Azure's biggest AI customer. In April 2026, it renegotiated that partnership, ending exclusivity and revenue-sharing arrangements. What followed is a family of seven models built entirely in-house, with no distillation from third-party models.

The Microsoft MAI models aren't just a product refresh. They represent the first time enterprises evaluating AI have a credible Microsoft-native option that doesn't route through OpenAI.

Teams comparing foundation models today are evaluating far more than raw benchmark scores. Cost, context windows, tool use, and deployment requirements all factor into selecting the right model for AI applications.

Microsoft MAI models improve what the model can do. Agent Builder changes what it outputs. Build agents that respond with interactive charts and tables instead of text.

Try Agent Builder Today

Microsoft MAI Models: The Full Family

The Microsoft MAI Models include seven models across reasoning, coding, image generation, transcription, and voice:

Model	Category	Status
MAI-Thinking-1	Reasoning	Private preview
MAI-Code-1-Flash	Coding	Generally available
MAI-Image-2.5	Image generation	Available
MAI-Image-2.5-Flash	Image generation (fast)	Available
MAI-Transcribe-1.5	Audio transcription	Available
MAI-Voice-2	Voice agents	Available
Voice-2-Flash	Voice agents (fast)	Available

The two models worth examining closely are MAI-Thinking-1 and MAI-Code-1-Flash. The others round out the portfolio, but these two carry the most weight for developers and enterprise teams making near-term decisions.

MAI-Thinking-1: Microsoft MAI Models' Flagship Reasoner

Architecture and Context

MAI-Thinking-1 is a sparse mixture-of-experts model with 35 billion active parameters and approximately 1 trillion total parameters. It supports a 256K context window. Microsoft describes it as built from scratch on commercially licensed data, with no distillation from other models.

Why does that matter? Enterprises that have blocked DeepSeek or other models due to IP provenance concerns get a clean alternative here. Microsoft's "zero distillation" claim, if it holds under scrutiny, removes a compliance obstacle that has slowed AI adoption in regulated industries.

Benchmark Performance

According to Microsoft's own numbers:

AIME 2025: 97%
SWE-Bench Pro: 53% (matches Claude Opus 4.6)

The AIME score places it among the strongest reasoning models publicly benchmarked. The SWE-Bench Pro result is notable because it comes from a model Microsoft claims is priced at roughly 10x better output tokens per dollar compared to GPT-5.5.

A note on verification: These figures come from Microsoft. Independent benchmarking hasn't caught up yet. They're worth tracking as third-party results emerge, but shouldn't be treated as settled.

Microsoft also reports that independent raters prefer MAI-Thinking-1 over Claude Sonnet 4.6 in blind side-by-sides. Again, this is Microsoft's own claim and hasn't been independently verified.

Availability

MAI-Thinking-1 is in private preview on Microsoft Foundry. Most teams can't access it today. If this model is central to your evaluation, expect to wait.

MAI-Code-1-Flash: The Microsoft MAI Models Coding Pick

Architecture

MAI-Code-1-Flash is a 5-billion-parameter model with a 256K context window. For context, that puts it in the same size class as Claude Haiku 4.5, not a heavyweight frontier model. The benchmark performance relative to its size is the story here.

What It Can Do

SWE-Bench Pro: 51% at 5B parameters
Solves harder problems with up to 60% fewer tokens than comparable models
Adaptive reasoning: It scales its compute to task complexity, staying concise on simple requests and spending more tokens on difficult ones

That 51% SWE-Bench Pro figure is meaningful because it approaches MAI-Thinking-1's 53% at a fraction of the compute cost. For high-volume coding pipelines where cost compounds at scale, this changes the math considerably.

Microsoft says it beats Claude Haiku 4.5 on price-to-performance across coding benchmarks.

Pricing

Input: $0.75 per million tokens
Output: $4.50 per million tokens

At this price point, it sits below most frontier models and competes directly with small-model alternatives from Anthropic and Google.

Availability

MAI-Code-1-Flash is live now in the GitHub Copilot model picker in VS Code, available across all tiers: Free, Pro ($10/month), and Pro+ ($39/month). It's accessible today, not gated behind a preview program.

One relevant detail: the model was trained inside GitHub Copilot's production harness, not evaluated against it after the fact. That's a different approach from most coding models, which are trained on general code and then benchmarked on tools they've never interacted with before.

Microsoft MAI Models’ Benchmark Summary

Model	Benchmark	Score	Context
MAI-Thinking-1	AIME 2025	97%	Microsoft-reported
MAI-Thinking-1	SWE-Bench Pro	53%	Matches Claude Opus 4.6
MAI-Code-1-Flash	SWE-Bench Pro	51%	At 5B parameters
MAI-Image-2.5	Arena Score	1403	#3 text-to-image, #2 image-to-image

The Broader Picture: Microsoft and OpenAI

The launch of Microsoft MAI Models coincides with a structural shift. Microsoft's renegotiated agreement with OpenAI ends the exclusivity and revenue-sharing that defined the partnership since 2019. Azure no longer has to route enterprise AI workloads through OpenAI models.

What this means in practice:

Microsoft can price its own models without coordinating with OpenAI on margins
Azure customers get native alternatives that don't carry OpenAI's IP provenance questions
GitHub Copilot can now default to MAI-Code-1-Flash without licensing implications

This isn't a clean break. Microsoft still distributes OpenAI models through Azure and maintains its investment. But for the first time, the relationship is competitive in parts of the stack that it wasn't before.

How Microsoft MAI Models Compare to the Competition

A brief positioning note rather than a full comparison:

vs. GPT-5.5: MAI-Thinking-1 claims roughly 10x better output token efficiency at comparable task performance. GPT-5.5 has broader third-party validation.
vs. Claude Opus 4.8: MAI-Thinking-1 matches Claude Opus 4.6 on SWE-Bench Pro. The Opus 4.8 comparison is pending independent benchmarks.
vs. Gemini 3.1 Pro: No direct public comparison available yet.
vs. Cursor Composer 2.5 and AI coding assistants: MAI-Code-1-Flash's integration directly into GitHub Copilot, trained on production harness data, gives it a distribution advantage over tools that rely on third-party model APIs.

Build Agents That Go Beyond Text

MAI-Thinking-1 and MAI-Code-1-Flash improve the underlying model. Agent Builder by Thesys focuses on the experience layer, helping teams create AI agents that respond with interactive charts, tables, forms, and dynamic visual interfaces instead of plain text.

Build Your First AI Agent Today

FAQs: Microsoft MAI Models

What are Microsoft MAI models?

The MAI models are a family of seven AI models built by Microsoft without distillation from third-party models. They cover reasoning, coding, image generation, transcription, and voice. The family was announced following Microsoft's renegotiation of its OpenAI partnership in April 2026.

What is MAI-Thinking-1?

MAI-Thinking-1 is Microsoft's flagship reasoning model. It uses a sparse mixture-of-experts architecture with 35 billion active parameters and approximately 1 trillion total parameters, with a 256K context window. It's currently in private preview on Microsoft Foundry and not yet publicly accessible.

What is MAI-Code-1-Flash?

MAI-Code-1-Flash is a 5-billion-parameter coding model with a 256K context window. It's available now in the GitHub Copilot model picker in VS Code across all subscription tiers. It achieves 51% on SWE-Bench Pro and is priced at $0.75 per million input tokens and $4.50 per million output tokens.

How does MAI-Thinking-1 compare to Claude Opus 4.8 and GPT-5.5?

According to Microsoft's own benchmarks, MAI-Thinking-1 scores 97% on AIME 2025 and 53% on SWE-Bench Pro. Microsoft claims it's priced at approximately 10x better output token efficiency than GPT-5.5. A direct comparison with Claude Opus 4.8 on current benchmarks hasn't been published yet by independent evaluators.

Is MAI-Code-1-Flash available now?

Yes. MAI-Code-1-Flash is available in the GitHub Copilot model picker in VS Code today, on Free, Pro ($10/month), and Pro+ ($39/month) tiers.

What is the pricing for MAI models?

MAI-Code-1-Flash is priced at $0.75 per million input tokens and $4.50 per million output tokens. Pricing for MAI-Thinking-1 has not been publicly announced, as it remains in private preview.

How are Microsoft MAI models different from OpenAI models?

Microsoft built the MAI models in-house without distilling from OpenAI or other third-party models. They run on Azure infrastructure but are independent of OpenAI's model weights, training data, and pricing. This distinction matters for enterprises with IP compliance requirements, since there are no provenance questions tied to other models' training data.

Does MAI-Thinking-1 use OpenAI training data?

Microsoft states that MAI-Thinking-1 was built from scratch on commercially licensed data with zero distillation from third-party models, including OpenAI's. This claim hasn't been independently verified yet, but it's the basis of Microsoft's commercially safe positioning for enterprise customers.

Related articles

May 2026 Security Incident

Why We're Open Sourcing OpenUI

AI Copilots: A Complete Guide to Intelligent Assistance

What Are AI Agents: Types, Capabilities, and Real-World Use

ChatKit: What It Is, What It Isn’t, and Where C1 Fits In

PowerPoint Automation Reimagined with C1 Artifacts API

The Future of AI Interfaces is Generative

MCP UI: Making AI Apps Interactive

What Is AG-UI and Why It Matters for AI-Native User Interfaces

How to design AI-Native Conversational Interfaces : From Templates to Generative UI

GPT 5 vs. GPT 4.1

How to Build Generative UI Applications

Implementing Generative Analytics with Thesys and MCP

Evolution of Analytics: From Static Dashboards to Generative UI

Why Generating Code for Generative UI is a bad idea

Building the First Generative UI API: Technical Architecture and Design Decisions Behind C1

How we evaluate LLMs for Generative UI

Generative UI vs Prompt to UI vs Prompt to Design

Generative UI: What It Is and Why It Matters