Microsoft MAI-Transcribe-2 Targets OpenAI and Google at $0.10 an Hour
Microsoft's MAI-Transcribe-2 offers 60-language speech recognition at $0.10 an hour, claiming benchmark-leading accuracy and sharply faster inference.
Summary
Microsoft AI released MAI-Transcribe-2 on September 3, 2026, through Microsoft Foundry and MAI Playground. Its introductory price is $0.10 per audio hour, 72% below the $0.36 charged for April’s MAI-Transcribe-1, cutting 100,000 annual hours from $36,000 to $10,000. Coverage has risen from 25 languages on April 2 to 43 with MAI-Transcribe-1.5 on June 2 and 60 now. The base rate includes speaker diarization, word timestamps, keyword biasing, automatic language detection, verbatim and cleaned output, and mid-sentence code switching, including Hinglish and Spanglish.
Microsoft claims a leading 5.2% average word error rate across 60 languages on FLEURS, although the read-speech benchmark does not represent conversation and MAI-Transcribe-1.5 scored 3.7% across 43 languages. Artificial Analysis ranks the new model second for accuracy and on its accuracy-latency Pareto frontier. Microsoft says it is 10 times faster than OpenAI’s GPT-Transcribe, seven times faster than ElevenLabs’ Scribe v2, and five times faster than Google’s Gemini 3.5 Transcribe. Alibaba remains the accuracy rival Microsoft does not claim to beat.
The third release in five months advances Microsoft’s effort to replace external AI across Teams, Nuance, Azure, Word and Excel after investing more than $13 billion in OpenAI. Mustafa Suleyman credited the original model to a focused 10-person team and said it used half the GPU cost of comparable leading models. Buyers still lack the introductory offer’s end date and standard price, streaming performance, per-language accuracy, diarization error rates, and policies covering residency, retention and training use of submitted audio.
Positives
- $0.10 per audio hour cuts MAI transcription pricing by roughly 72% from April’s $0.36 rate.
- 60 supported languages more than double the 25 available with MAI-Transcribe-1 on April 2.
- Speaker diarization, timestamps, keyword biasing, language detection, output controls and code switching are included in the base price.
- Artificial Analysis places MAI-Transcribe-2 second for accuracy and on the accuracy-latency Pareto frontier.
- Five to 10 times faster claimed inference could reduce GPU usage and improve high-volume batch throughput.
- 100,000 annual transcription hours would cost $10,000 instead of $36,000 at the introductory rate.
Risks & concerns
- The $0.10 introductory offer has no published end date or confirmed standard rate.
- A 5.2% FLEURS average conceals per-language variation and trails MAI-Transcribe-1.5’s 3.7% across its narrower 43-language set.
- FLEURS measures read speech, so it does not establish performance on noisy, overlapping business conversations.
- Microsoft disclosed no real-time streaming results, limiting evaluation for voice agents and live captions.
- No diarization error rate shows whether the model attributes otherwise accurate words to the correct speakers.
- Data residency, retention and possible training use remain unspecified for sensitive medical, legal and financial audio.