MAI (Microsoft AI) is the family of foundation models Microsoft develops itself. They are hosted and billed through Azure's Microsoft Foundry (formerly Azure AI Foundry) and distributed Azure-first. They sit alongside OpenAI models (via Azure OpenAI) as a separate family that Microsoft develops and serves itself. Specific prices are shown in the table below; the goal is to give enterprises more choice on cost, licensing and data sovereignty.
Latest models (October 2026)
The most recent launches confirmed on the microsoft.ai blog, including the October 2, 2026 release. All are available via Microsoft Foundry / MAI Playground and other routes.
🎧 MAI-Transcribe-2-Streaming (streaming speech recognition) — Oct 2, 2026
MAI's first streaming speech recognition model. Instead of waiting for the speaker to finish, it returns its first partial result (a provisional transcript produced mid-speech) in just over 100ms of receiving audio, covers 60 languages, and supports automatic continuous language detection. Access is via Microsoft Foundry / MAI Playground / OpenRouter / Vercel / Azure Voice Live (LiveKit is "coming soon"). The same announcement introduces the speech-generation models MAI-Voice-2.1 / 2.1-Flash (23 languages, 26 locales). Prices are collected in the "Reference prices" table below. By "verified pricing table" we mean the Azure official pricing page that this site checks directly — as of 2026-10-17 neither this model nor the Voice-2.1 family appears in it.
Verbatim from the official blog: "MAI-Transcribe-2-Streaming delivers low-latency, real-time transcripts in 60 languages, all while supporting automatic, continuous language detection." / "Rather than waiting for someone to finish speaking before returning text, it produces its first hypotheses (known as "partials") in just over 100ms of receiving audio." / "MAI-Transcribe-2-Streaming is available at an introductory price of $0.54 per hour of audio through the end of the year."
🖼️ MAI-Image-2.6 / MAI-Image-2.6-Flash (image generation) — Sep 4, 2026
Microsoft's strongest image models. They add multi-reference image editing (compose people, products, styles and scenes from multiple images), web grounding, dynamic aspect ratios, and support up to 1.5K resolution. MAI-Image-2.6 targets maximum precision/quality; MAI-Image-2.6-Flash delivers comparable quality for latency-sensitive, high-throughput workloads. Per Microsoft's announcement (as of Sep 4, 2026), MAI-Image-2.6 is No. 2 on Arena for both text-to-image and image editing, and No. 1 on Artificial Analysis for image editing. Microsoft reports 2.6-Flash is 2.8× faster and 72% more efficient than GPT-Image-2-Medium. The blog lists $5 text input / $8 image input / $38 image output for MAI-Image-2.6 and $1.75 / $2.50 / $19 for 2.6-Flash (starting prices as stated on the blog). Both are in Public Preview.
🎙️ MAI-Transcribe-2 (speech recognition) — Sep 3, 2026
Microsoft's blog title positions it as the "fastest, most accurate and cheapest speech recognition model in the world." Priced at $0.10 / hour (a limited-time offer through the end of the year). It is No. 1 on FLEURS across 60 languages (avg WER 5.2%) and No. 2 on Artificial Analysis's AA-WER leaderboard (2.0%). Microsoft cites speeds (via Artificial Analysis) of 10× GPT-Transcribe, 7× Scribe v2, and 5× Gemini 3.5 Transcribe. Features include speaker diarization, word-level timestamps, keyword biasing, verbatim/clean transcription styles, code switching and automatic language identification.
Sources: Pushing the quality-cost frontier with MAI-Image-2.6 (2026-09-04) · MAI-Transcribe-2 (2026-09-03) · Our first streaming transcription model (2026-10-02)
How to read the table (image generation vs image understanding): in the Microsoft section of Azure Foundry Models pricing, the Input column differs by model. MAI-Image-2.5 / 2.5 Flash / 2.5 Pro list two input kinds, Text and Image, with separate rates for passing an image as input. MAI-Image-2 / MAI-Image-2-Efficient list text input only, and text models such as MAI-Thinking-1 also list text input only. In short: a split Text / Image input row means the model accepts images (understanding), while an image in the Output column means the model generates images. (Re-checked on the official pricing page on 2026-10-17.)
How to read the table (image generation vs image understanding): in the Microsoft section of Azure Foundry Models pricing, the Input column differs by model. MAI-Image-2.5 / 2.5 Flash / 2.5 Pro list two input kinds, Text and Image, with separate rates for passing an image as input. MAI-Image-2 / MAI-Image-2-Efficient list text input only, and text models such as MAI-Thinking-1 also list text input only. In short: a split Text / Image input row means the model accepts images (understanding), while an image in the Output column means the model generates images. (Re-checked on the official pricing page on 2026-10-17.)
Reference prices: MAI-Transcribe-2-Streaming / MAI-Voice-2.1 / 2.1-Flash (from the official blog; not in the pricing table)
These three models are also not in the verified table above (Azure Foundry Models pricing, Microsoft section) as of 2026-10-03: the Microsoft section still has nine MAI rows, with no Transcribe or Voice rows. So we publish the figures exactly as stated in the official post (2026-10-02).
| Model | Type | As stated on the official blog |
|---|---|---|
| MAI-Transcribe-2-Streaming | Streaming speech recognition | $0.54 / hour of audio (introductory, through 2026) |
| MAI-Voice-2.1 | Speech generation | $22 / 1M characters |
| MAI-Voice-2.1-Flash | Speech generation (fast) | $15 / 1M characters |
Model pricing (Azure Foundry, verified)
From the Microsoft Azure "Foundry Models pricing (Microsoft section)" in US dollars. Prices are USD per 1M tokens; image and speech models use special units — see notes. Verified 2026-09-09 (re-checked 2026-10-03: the 2.6 family, Transcribe-2-Streaming and the Voice-2.1 family are all still absent from that table, and existing rows are unchanged).
| Model | Input / 1M | Cached Input / 1M | Output / 1M | Category |
|---|---|---|---|---|
| MAI-Thinking-1 | $2.00 | $0.20 | $8.00 | Reasoning |
| MAI-Cyber-1 Flash | $0.60 | $0.06 | $3.50 | Cyber security (MDASH) |
| MAI-DS-R1 | $1.35 | — | $5.40 | DeepSeek R1 family (Regional $1.485/$5.94) |
| MAI-Image-2 | $5 | — | $33 | Image generation (Global) Image out |
| MAI-Image-2-Efficient | $5 | — | $19.50 | Image generation (41% lower cost) Image out |
| MAI-Image-2.5 | Text $5 / Image $8 | — | $47 | Image generation Image out |
| MAI-Image-2.5 Flash | Text $1.75 / Image $1.75 | — | $19.50 | Image generation (fast) Image out |
| MAI-Image-2.5 Pro | Text $5 / Image $8 | — | $106 | Image generation (high quality) Image out |
Reference prices: MAI-Image-2.6 / 2.6-Flash (from the official blog; not in the pricing table)
These two models are not in the verified table above (Azure Foundry Models pricing, Microsoft section) as of 2026-10-03. So we publish the prices exactly as stated on the Microsoft Foundry blog post (2026-09-04). All figures are starting prices and both models are in public preview.
| Model | Text input / 1M | Image input / 1M | Image output / 1M |
|---|---|---|---|
| MAI-Image-2.6 | $5 | $8 | $38 |
| MAI-Image-2.6-Flash | $1.75 | $2.50 | $19 |
Text input and image input are billed at different rates (image input costs more), and the output rate applies to image output. Against MAI-Image-2.5 in the verified table ($5 / $8 / $47), MAI-Image-2.6 comes in at $38 output and 2.6-Flash at $1.75 / $2.50 / $19.
Verbatim from the official blog (pricing paragraph): "MAI-Image-2.6 and MAI-Image-2.6-Flash are available in public preview in Foundry. MAI-Image-2.6 pricing starts at $5 USD per 1M tokens for text input, $8 USD per 1M tokens for image input, and $38 USD per 1M tokens for image output. MAI-Image-2.6-Flash pricing starts at $1.75 USD per 1M tokens for text input, $2.50 USD per 1M tokens for image input, and $19 USD per 1M tokens for image output."
- MAI-Image-2.6 / 2.6-Flash are still not in the verified Microsoft Foundry pricing table, so the figures above come from the official blog (starting prices, public preview). In public preview, pricing and availability can move — confirm the current numbers in Foundry before contracting.
- MAI-Transcribe-2-Streaming / MAI-Voice-2.1 / 2.1-Flash are likewise absent from the verified pricing table, so the figures above come from the official blog. Note that the $0.54/hour for MAI-Transcribe-2-Streaming is explicitly an introductory price through the end of the year.
- MAI-Code family (MAI-Code-1-Flash / 1.1-Flash) are also absent from that table, so pricing is omitted.
- MAI-Voice family (MAI-Voice-1 / 2 / 2-Flash) are officially served via Azure Speech in Foundry Tools and billed via speech meters, not tokens (so not in this table).
Primary source: azure.microsoft.com/.../ai-foundry-models/microsoft/ (verified 2026-10-03) · MAI-Image-2.6 / 2.6-Flash pricing only: Microsoft Foundry Blog, MAI-Image-2.6 and MAI-Image-2.6-Flash (2026-09-04; verified 2026-10-03)
How to use (API / access)
MAI models are primarily deployed and called via Azure AI Foundry / Microsoft Foundry. Some can also be tried in MAI Playground or via OpenRouter.
- Microsoft Foundry (ai.azure.com/catalog/models): pick a MAI model from the catalog and deploy it by model ID. Auth is Azure (API key / Entra ID, similar to Azure OpenAI).
- MAI Playground (playground.microsoft.ai): try models in the browser without an API key. Examples:
playground.microsoft.ai/chat?model=mai-image-2-6-flash/?model=mai-transcribe-2. - MAI-Transcribe-2 (speech): officially via Azure Speech MAI-Transcribe docs (REST / SDK). On OpenRouter the model ID is
microsoft/mai-transcribe-2. - MAI-Transcribe-2-Streaming (Oct 2, 2026; streaming): the official post lists OpenRouter / Microsoft Foundry / MAI Playground / Vercel / Azure Voice Live (LiveKit is "coming soon"). It targets cases where waiting for the speaker to stop is not an option, such as live meeting transcription.
- Image generation (MAI-Image-2.x): deploy as a text-to-image / image-editing model in Foundry. Multi-reference editing and web grounding are supported by the MAI-Image-2.6 generation.
- Reasoning (MAI-Thinking-1): a reasoning model; note the cheap cached-input rate of $0.20.
⚠️ Exact request formats and parameters differ by model and may change, so always consult the official documentation (Learn / the specific Foundry model page). This guide intentionally does not include concrete API request bodies, as those were not verified against a primary source.
MAI release timeline
Dates from the microsoft.ai blog, key highlights.
| Date | Milestone |
|---|---|
| ~Sep 2025 | “Two in-house models” and “Introducing MAI-Image-1” (MAI-Image-1 debuts in LMArena top 10) |
| 2026-04-02 | “Announcing 3 new world class MAI models” = MAI-Image-2 / MAI-Transcribe-1 / MAI-Voice-1 (available in Foundry) |
| 2026-04-02 | MAI-Image-2-Efficient (flagship quality, 41% lower cost) |
| May 2026 | MAI-Transcribe-1.5 / MAI-Code-1-Flash announced |
| 2026-06-02 | Build 2026 “Building a hill-climbing machine: Launching seven new MAI models” = MAI-Image-2.5 / 2.5-Flash / Transcribe-1.5 / MAI-Thinking-1 / Voice-2 / Voice-2-Flash / Code-1-Flash |
| Jul 2026 | MAI-Cyber-1-Flash (MDASH, security) / MAI-Image-2.5-Pro & MAI-Voice-2-Flash / MAI-Image-2.5 (No. 2 on Arena) |
| Aug 2026 | MAI-Code-1.1-Flash (at a quarter of the cost) / MAI-Image-2.6 reaches No. 2 on Arena |
| 2026-09-03 New | MAI-Transcribe-2 (speech recognition, $0.10/hr limited) |
| 2026-09-04 New | MAI-Image-2.6 / 2.6-Flash (image generation) |
| 2026-10-02 New | MAI-Transcribe-2-Streaming (streaming speech recognition, $0.54/hr introductory through 2026) plus MAI-Voice-2.1 / 2.1-Flash (speech generation) announced |
| 2026-10-07 New | A local version of MAI Code 1.1 Flash announced on Microsoft’s developer blog, running on-device in GitHub Copilot on Windows (137B total / 6.8B active MoE; the quantized build is about 53GB and peaks at 75.5GB of memory at 256k context) |
Source: microsoft.ai/blog · commandline.microsoft.com (Microsoft developer blog)
Key things to know
- First-party models: MAI is a separate lineage from OpenAI models (served via Azure OpenAI). It is offered to enterprises as a way to broaden options for data sovereignty and licensing.
- Azure-first: MAI models ship alongside open models from Google, Meta, Mistral, DeepSeek and others in the Foundry catalog (11,000+ models), using unified Azure billing and operations.
- Price examples: The reasoning model MAI-Thinking-1 is $2/$8 (cached input $0.20); image generation, e.g. MAI-Image-2.5 Flash, is $1.75/$19.50. The newest MAI-Image-2.6 is $5 text input / $8 image input / $38 image output, and 2.6-Flash is $1.75 / $2.50 / $19 (all starting prices, public preview, from the official blog).
- Newest speech recognition (2026-10-02): MAI-Transcribe-2-Streaming transcribes 60 languages in streaming mode and returns a partial in about 100ms without waiting for the speaker to stop. Introductory price through the end of the year: $0.54/hour (official blog). It is not yet in the verified Azure pricing table.
- Local (on-device) push (2026-10-07) — memory needs, supported hardware and how to enable it: Microsoft’s developer blog announced a local version of MAI Code 1.1 Flash. It is a coding-optimized MoE with 137B total and 6.8B active parameters; the on-device build uses quantization and speculative decoding to shrink to about 53GB (an 80% reduction versus the Bfloat16 cloud variant) and peaks at 75.5GB of memory at 256k context (tested 2026-10-05). It ships first in GitHub Copilot on NVIDIA RTX Spark Windows PCs (Surface Laptop Ultra), sandboxed by the Windows team’s open-source Microsoft Execution Containers (MXC). There are two ways to use it across the GitHub Copilot CLI, Copilot app and VS Code: select MAI Code 1.1 Flash through the Windows ML provider, or connect Copilot to OpenAI-compatible local endpoints. The post lists no pricing and MAI-Code-1.1 is absent from the Azure pricing table as of 2026-10-13, so it cannot be used for cost comparisons yet.
- Broad modality coverage: Speech recognition (Transcribe), speech generation (Voice), image generation (Image), image understanding (VLM), coding (Code) and reasoning (Thinking) extend well beyond plain LLM text.
⚠️ Disclaimer
- Pricing and specs were checked against Microsoft's official pages on 2026-10-03 (the verified pricing table was first checked on 2026-09-09, re-checked 2026-09-19 and 2026-10-03) and are subject to change. Always verify on the official pricing page and official blog before contracting.
- Benchmarks (Arena / Artificial Analysis rankings, FLEURS WER, speed multiples, 2.8× faster etc.) are self-reported by Microsoft / Artificial Analysis and may diverge from independent evaluation.
- MAI-Image-2.6 / 2.6-Flash pricing comes from the official blog (starting prices, public preview), not from the verified pricing table. Prices for the MAI-Code and MAI-Voice families are omitted because no primary source figure exists. MAI-Transcribe-2-Streaming / MAI-Voice-2.1 / 2.1-Flash figures are likewise from the official blog (introductory pricing) and are not in the verified pricing table.
- This site is informational only; we are not affiliated with or endorsed by any provider. Microsoft and Azure are trademarks of Microsoft Corporation; MAI is a trademark of Microsoft AI.