MAI-Transcribe-2 is Microsoft’s new speech-to-text model, and it arrives with a headline price that is hard to ignore: $0.10 per hour of audio. Microsoft AI released it on September 3, 2026, in public preview through Microsoft Foundry, and it is also available in the MAI Playground and on OpenRouter.
Microsoft calls MAI-Transcribe-2 the fastest, most accurate and cheapest speech recognition model in the world. Some of that claim holds up in independent testing, and some of it depends on which numbers you look at.
This guide separates the confirmed facts from the marketing, and explains what the launch means for small businesses that transcribe meetings, sales calls, interviews or podcasts.
Table of Contents
What Is MAI-Transcribe-2?
MAI-Transcribe-2 is the successor to MAI-Transcribe-1.5 and part of Microsoft AI’s in-house MAI model family. It converts recorded audio into text and supports 60 languages, up from 43 in the previous version.
It identifies the language automatically, handles recordings that mix several languages, and follows speakers who switch languages mid-sentence (code-switching). Speaker diarization (labelling who said what), word-level timestamps and keyword biasing are built in.
MAI-Transcribe-2 Pricing: $0.10 per Audio Hour (For Now)
The launch price is $0.10 per hour of audio. That is roughly a 72% cut from the $0.36 per hour Microsoft charged for its previous transcription model a few months earlier.
There is an important catch: Microsoft describes the $0.10 rate as a limited-time offer that runs until the end of 2026, and it has not said what the model will cost from January. Treat the current price as an introductory rate, not a long-term budget figure.
To put the numbers in context, a business transcribing 100,000 hours of call-centre audio a year would pay about $10,000 at the launch rate, compared with about $36,000 at the earlier price. A small team transcribing 20 hours of meetings a month would pay around $2.
Accuracy: How MAI-Transcribe-2 Compares
Microsoft’s own figures use the FLEURS multilingual benchmark. MAI-Transcribe-2 posts an average word error rate (WER) of 5.2% across its 60 languages, and 3.4% across the top 25 languages, compared with 3.7% for MAI-Transcribe-1.5.
Microsoft says it beats Whisper-Large-V3, GPT-Transcribe, ElevenLabs Scribe v2 and Gemini 3.5 transcription on that benchmark. Those are vendor-reported results, so the independent numbers matter more.
On Artificial Analysis, MAI-Transcribe-2 scores a 2.0% AA-WER, which ranks it second on the leaderboard, slightly ahead of ElevenLabs Scribe v2 at 2.2%.
| Metric | MAI-Transcribe-2 |
|---|---|
| Price | $0.10 per audio hour (launch offer until end of 2026) |
| Languages | 60, with automatic language identification |
| FLEURS WER (Microsoft) | 5.2% average; 3.4% on the top 25 languages |
| AA-WER (Artificial Analysis) | 2.0%, second on the leaderboard |
| Speed | About 10 seconds of inference per hour of audio |
| Availability | Microsoft Foundry (public preview), MAI Playground, OpenRouter |
Word error rate varies a lot with audio quality, accents and vocabulary. Run a test on your own recordings before you switch providers.
Speed and Features
Speed is where the model stands out most. Microsoft says one hour of audio takes about 10 seconds of model inference, and Artificial Analysis measures a median speed factor of roughly 410x real time (other sources cite about 404x).
It is not the fastest model on raw throughput: Deepgram Nova-3 runs at around 600x, but with a materially higher error rate. For most business use, the balance of speed and accuracy matters more than throughput alone.
Built-in features include:
- Automatic language identification across 60 languages
- Multilingual recordings and code-switching
- Speaker diarization
- Word-level timestamps
- Keyword biasing, to improve accuracy on product names and industry terms
Where You Can Use It Today
At launch, the model is available in three places: Microsoft Foundry (in public preview), the MAI Playground for quick tests, and OpenRouter. If you already route model calls through OpenRouter, trying it is mostly a matter of changing the model name.
Microsoft has not announced it as a built-in feature of Teams or Microsoft 365 plans, so do not assume an existing subscription includes it. Check the Foundry pricing page before you commit.
What MAI-Transcribe-2 Means for Small Businesses
For small businesses, the practical impact is cost. At $0.10 per audio hour, MAI-Transcribe-2 makes it realistic to transcribe every sales call, support call or team meeting, not just the important ones.
Common workflows include:
- Sales calls into searchable transcripts in your CRM
- Meetings into summaries and action items with a language model
- Podcasts and videos into blog posts, newsletters and social snippets
- Support calls into reports on recurring questions and complaints
Transcription is only the first step. Pairing it with an automation tool and a language model is what turns raw text into summaries, CRM updates and content. Our guides on AI agents for small business and AI for content creation cover the next steps.
Limitations and Open Questions
- Price after 2026: unknown. Budget for a higher rate from January.
- Preview status: it is a public preview on Foundry, so features and limits may change.
- Vendor benchmarks: the FLEURS figures are self-reported; the Artificial Analysis results are the independent check.
- Compliance: if you handle health, legal or financial recordings, confirm data handling and certifications with Microsoft for your region before sending audio.
The Bottom Line
MAI-Transcribe-2 is a strong release: independently ranked among the most accurate transcription models, very fast, and, for now, extremely cheap. For small businesses it is worth testing on real recordings while the $0.10 launch price lasts.
Plan for two things: the price will likely change in January, and the preview may change before general availability. Keep your pipeline model-agnostic so you can switch transcription providers if the economics change.