What are the seven Microsoft MAI models?

At Microsoft Build 2026 on June 2, 2026, Microsoft announced seven in-house MAI (Microsoft AI) models: MAI-Thinking-1 (frontier reasoning), MAI-Code-1-Flash (coding model), MAI-Image-2 (image generation), MAI-Voice-1 (text-to-speech), MAI-Transcribe-1 (speech-to-text), and two additional smaller models for specific Foundry workloads. The flagship is MAI-Thinking-1 (97% AIME 2025).

Is Microsoft replacing OpenAI in Copilot?

Not entirely, but it's clearly reducing dependence. Microsoft positioned MAI as part of a 'long-term self-sufficiency' strategy. Copilot still uses OpenAI models heavily, but Microsoft is now routing growing portions of Copilot traffic to MAI models — especially for tasks where MAI is competitive. The Microsoft-OpenAI partnership runs through 2030 but with looser exclusivity.

How does MAI-Thinking-1 compare to GPT-5.5?

On pure math, MAI-Thinking-1 leads — 97.0% AIME 2025 vs GPT-5.5's ~94%. On agentic coding, Microsoft says MAI-Thinking-1 matches Claude Opus 4.6 on SWE-Bench Pro, putting it near GPT-5.5 thinking-mode performance. GPT-5.5 still wins on multimodal breadth (vision/audio) and the broader plugin ecosystem. MAI-Thinking-1 is expected to be roughly half the price of GPT-5.5 thinking in Azure Foundry.

Was MAI-Thinking-1 trained without OpenAI data?

Yes. Microsoft explicitly stated that MAI-Thinking-1 was trained from scratch without using OpenAI outputs or training data, which is significant because of the OpenAI exclusivity clauses in their existing partnership. The training reportedly used Microsoft's own compute and curated datasets on H200 and Maia 3 silicon.

Quick Answer

What Is Microsoft's MAI Family? Seven Models, OpenAI Independence

Published: June 4, 2026

What Is Microsoft’s MAI Family? Seven Models, OpenAI Independence

Microsoft just announced seven in-house AI models at Build 2026 — a clear signal that the OpenAI honeymoon is ending. Led by MAI-Thinking-1 (97% AIME 2025), the MAI family is Microsoft’s bid for AI independence.

Last verified: June 4, 2026

The seven models

Model	Type	Highlight
MAI-Thinking-1	Frontier reasoning	97.0% AIME 2025; matches Claude Opus 4.6 on SWE-Bench Pro
MAI-Code-1-Flash	Fast coding model	Sub-second responses, in-IDE Copilot routing
MAI-Image-2	Image generation	Native 2K resolution, transparent backgrounds
MAI-Voice-1	Text-to-speech	Emotive multi-voice synthesis, sub-300ms latency
MAI-Transcribe-1	Speech-to-text	Real-time multilingual transcription, beats Whisper v4 on many languages
MAI-Vision-1	Image understanding	OCR + visual reasoning
MAI-Embed-1	Embeddings	Domain-tuned for enterprise document retrieval

All seven are available through Azure AI Foundry, with the largest models also integrated into Microsoft 365 Copilot, GitHub Copilot, and Windows.

Why now

Microsoft and OpenAI have been the defining AI partnership since 2019. So why is Microsoft suddenly running its own race?

1. Dependency risk

OpenAI’s exclusivity clauses, IPO ambitions, and compute demands made Microsoft a hostage to a single supplier. By 2026, every major Copilot product line ran on OpenAI infrastructure.

2. Cost

Inference cost on top-tier OpenAI models was eating into Copilot margins. MAI-Code-1-Flash and MAI-Thinking-1 give Microsoft cheaper alternatives for high-volume Copilot routing.

3. Strategic optionality

If OpenAI raises prices, exits the partnership, or has a safety incident, Microsoft can route Copilot traffic to MAI without disrupting users.

4. Talent and infrastructure

After hiring Mustafa Suleyman (DeepMind co-founder, ex-Inflection CEO) as Microsoft AI CEO, Microsoft accumulated the team and compute to compete at the frontier.

What “self-sufficiency” actually means

Microsoft’s exact phrasing at Build 2026 was “long-term self-sufficiency.” That implies:

MAI handles the majority of internal AI workloads by some target year
OpenAI partnership continues for premium consumer/enterprise tiers and joint research
Customers can choose between MAI, OpenAI, Anthropic, Meta, and open models in Foundry
Copilot routes to whichever model is cheapest/best for each request

This is not Microsoft divorcing OpenAI. It’s Microsoft removing the single-vendor risk.

How this changes the AI market

For OpenAI

Reduces the most reliable single buyer of OpenAI capacity. OpenAI’s IPO narrative still works, but the “Microsoft is locked in” thesis weakens.

For Anthropic

Confirms Microsoft is increasingly multi-vendor. Anthropic’s Foundry availability is more important now — and may grow if MAI underperforms.

For Foundry developers

More choice, more price pressure, more transparent routing. Developers can mix MAI for cheap reasoning, GPT-5.5 for multimodal, Claude Opus 4.8 for coding — all within one Azure billing.

For OpenAI’s IPO

The S-1 will need to address whether Microsoft is still a sustaining customer at OpenAI’s projected scale. Expect tough questions.

Quick benchmark snapshot

Benchmark	MAI-Thinking-1	GPT-5.5	Claude Opus 4.8
AIME 2025	97.0%	~94%	~95%
AIME 2026	94.5%	~91%	~93%
SWE-Bench Pro	≈66% (≈Opus 4.6)	~67%	69.2%
Multimodal Vision	Limited	✅	✅

What developers should do

Try MAI-Thinking-1 in Foundry for cost-sensitive reasoning tasks
Add MAI-Code-1-Flash to Copilot routing for high-volume completions
Keep multi-model logic in production — don’t single-thread on any provider
Watch pricing — MAI is likely to be aggressively priced through 2026

Bottom line

The MAI family isn’t a research project — it’s a strategic pillar. With seven shipping models and a credible frontier reasoning entry, Microsoft has decoupled its product roadmap from OpenAI’s. Expect Copilot to feel increasingly Microsoft-native, with OpenAI showing up in fewer high-volume code paths over the next 12 months.