🇺🇸 USA · Microsoft MAI (Microsoft AI’s own foundation models)

Status: 🟩 COMPLETE 🟦 LIVING Last updated: 2026-06-26 Plain-English tagline: Microsoft’s own family of foundation models — MAI-1, MAI-Voice, MAI-Vision — built by the new Microsoft AI division under Mustafa Suleyman. The strategic effort to make Copilot work without depending only on OpenAI.


Front-matter facts

FieldValue
VendorMicrosoft AI (a Microsoft division) — led by Mustafa Suleyman (formerly co-founder of DeepMind, then Inflection AI)
Country / origin🇺🇸 USA (with UK heritage via Suleyman + many staff hired from Inflection / DeepMind)
Recommended for Australian users?✅ Yes — MAI models surface through Microsoft Copilot and Azure AI products in AUS
Privacy summaryConsumer Copilot uses MAI models per Copilot privacy posture; M365 Copilot / Azure no-train commitments apply
Free tierMAI models power free Copilot tier; not separately licensed to consumers
Paid tiersBundled into Copilot subscriptions (Pro, M365 Copilot, etc.); Azure AI Foundry for developers
First releasedMAI division formed March 2024 (when Microsoft hired Mustafa Suleyman + most of Inflection); first MAI-1 launched late 2024 / 2025
Last reviewed2026-06-26
Official sitehttps://www.microsoft.com/en-us/ai

What it is

Microsoft MAI = “Microsoft AI” — both the name of Microsoft’s AI division AND the name of its own foundation-model family (MAI-1, MAI-Voice, MAI-Vision, etc.). It’s the strategic effort by Microsoft to build its own frontier models internally, alongside (not replacing) its OpenAI partnership.

The backstory:

  • March 2024: Microsoft hired Mustafa Suleyman (co-founder of DeepMind, then CEO of Inflection AI) as CEO of a new “Microsoft AI” division. Most of Inflection’s research team came with him.
  • Inflection Pi (the consumer chatbot Suleyman built at Inflection) was effectively absorbed; the Inflection brand persists but the team rebuilt at Microsoft AI.
  • Why? Microsoft had been entirely dependent on OpenAI for frontier-AI capability — strategically risky. Building MAI gives Microsoft a backup, more control, and ability to differentiate Copilot.
  • MAI models have been gradually replacing OpenAI models in some Copilot features through 2024-26.

MAI’s published models / capabilities:

  • MAI-1 — first frontier-scale LLM from Microsoft AI; comparable to GPT-4 family at smaller sizes
  • MAI-Voice / MAI-Voice-1 — voice generation / TTS model
  • MAI-Vision — multimodal vision capabilities
  • MAI-1-Preview / MAI-1.5 — successor generations
  • Smaller / faster variants for on-device and edge use

MAI models are Microsoft-internal: they’re not licensable as a standalone API to the public the way Claude or GPT are. You access them indirectly via Microsoft Copilot, Microsoft 365 features, and (for developers) Azure AI Foundry.


What you’d use it for (indirectly)

You don’t “use MAI directly” — instead you use Microsoft products that MAI powers:

  • Microsoft Copilot (consumer, in Edge / Windows / Bing) — increasingly uses MAI models for some features
  • Microsoft 365 Copilot (work, in Office) — mixed OpenAI + MAI under the hood
  • Microsoft Designer — image gen via gpt-image-1 (OpenAI) but with MAI for some operations
  • GitHub Copilot — primarily OpenAI / Anthropic / Google models, but MAI has been added for some features
  • Azure AI Foundry — Microsoft’s enterprise AI platform; MAI models are listed alongside Claude / GPT / Llama / Mistral for developers
  • Cortana legacy — being replaced by MAI-powered Copilot

For developers using Azure, MAI models are an option alongside OpenAI / Anthropic / Llama in Azure AI Foundry — useful for first-party Microsoft compliance + the same Microsoft enterprise terms.


How to access from Australia

Indirectly (the main path)

  • Use Microsoft Copilot → some responses come from MAI models
  • Use Microsoft 365 Copilot at work → mixed OpenAI + MAI

Via Azure AI Foundry (for developers)

  1. Sign up at Azure (azure.microsoft.com)
  2. Open Azure AI Foundry
  3. Browse model catalog → MAI models listed alongside OpenAI / Claude / Llama / Mistral
  4. Same Azure billing and AUS data residency options (Australia East region)

What it costs

  • Bundled into Copilot subscriptions — no separate MAI consumer pricing
  • Azure AI Foundry — pay-per-token at Azure rates (verify current Azure pricing)
  • Enterprise — via Azure enterprise agreements with full data-residency / no-training options

How it compares to alternatives

AspectMicrosoft MAIOpenAI GPTAnthropic ClaudeGoogle Gemini
Primary accessVia Microsoft productsDirect (ChatGPT / API)Direct (Claude / API)Direct (Gemini / API)
Standalone subscriptionNo (bundled)Yes (ChatGPT)Yes (Claude.ai)Yes (Gemini)
Direct API for consumers / SMBsNo (Azure AI Foundry for enterprise)Yes (platform.openai.com)Yes (console.anthropic.com)Yes (Google AI Studio)
Frontier-scale capabilityApproaching frontierFrontierFrontierFrontier
MultimodalYes (MAI-Vision)YesYesYes
VoiceMAI-VoiceOpenAI Voice EngineRolling outGemini Live

MAI is strategically important for Microsoft (independence from OpenAI dependency) but functionally less important to most users than the frontier products it’s intermixed with.


Privacy / data handling

  • MAI models accessed via Microsoft Copilot / 365 Copilot / Azure inherit those products’ privacy postures
  • M365 Copilot / Azure AI Foundry: no training on customer data by default
  • Consumer Copilot: training defaults apply (opt-out available)
  • AUS data residency for enterprise via Azure Australia East

Recent changes

  • 2026: MAI-Vision and MAI-Voice integrations expanded across Copilot family
  • 2025: MAI-1 family models added to Azure AI Foundry; MAI-1.5 and successors
  • March 2024: Microsoft AI division formed under Mustafa Suleyman
  • March 2024: Microsoft hired most of Inflection AI’s team

Gotchas

  • You can’t easily “ask for MAI specifically” in consumer Copilot — Microsoft routes between MAI / OpenAI / others transparently. Some users find this opaque.
  • MAI ≠ OpenAI partnership replacement — Microsoft still works closely with OpenAI; MAI is additive
  • Inflection Pi as a product is largely defunct — the team moved to Microsoft; the Pi consumer chat has minimal updates since the transition
  • “Microsoft AI” division vs “Microsoft AI research” (the older AI Research arm) are different organisations — Suleyman’s division focuses on consumer-product-shipped AI
  • Frontier benchmarks — MAI models are competitive but as of mid-2026 not generally considered the frontier leader (GPT / Claude / Gemini hold that)
  • Azure AI Foundry is where developers actually use MAI — not a standalone-MAI website

See also


Sources