• Verdict: Wait. The number that decides it is the roughly 100ms to first partial, and Microsoft publishes no audio conditions, percentile or network path for it.
• Microsoft says the model ranks no. 1 for accuracy on both final and partial transcripts on Artificial Analysis, but its page links no methodology, and Artificial Analysis's own MAI-Transcribe-1 write-up does not state the AA-WER datasets either.
• At the introductory $0.54 per audio hour, the streaming model costs 1.5x the $6 per 1,000 minutes ($0.36 per hour) Artificial Analysis listed for MAI-Transcribe-1 on Foundry, though the two are different products.
• The $0.54 rate holds only through year-end; the standard rate was not stated on the page we read.
• Microsoft lists the model at 60 languages with continuous language detection, so test accented and code-switched audio, not just English.

Microsoft AI released MAI-Transcribe-2-Streaming, a real-time transcription model aimed at voice agents, alongside MAI-Voice-2.1 text-to-speech models. The announcement claims first partial transcripts in just over 100ms and a number-one accuracy ranking on Artificial Analysis. Our call is Wait: the only figure we could verify end to end is the price, and the headline latency comes with no stated test conditions.

The deciding number: just over 100ms to first partial

Microsoft says the model produces its first hypotheses, called partials, in just over 100ms of receiving audio. For a voice agent this sets how early downstream reasoning can start. But the announcement gives no percentile, audio type, chunk size or network path, so a reader cannot rerun it. A median on clean studio audio and a p95 on telephony audio are very different numbers, and the page does not say which this is.

The accuracy ranking cannot be reproduced from public text

Microsoft says the model ranks no. 1 for accuracy on both final and partial transcripts on Artificial Analysis, and that words appear 2x faster than its closest competitor, without naming that competitor. The ranking page is not linked from the announcement. Artificial Analysis's own MAI-Transcribe-1 write-up reports AA-WER scores, 3.0% and fourth place for that model, but does not describe the datasets or normalization behind the metric.

Price: 1.5x the earlier batch model, and only through year-end

Microsoft lists MAI-Transcribe-2-Streaming at $0.54 per audio hour as introductory pricing through the end of the year. Artificial Analysis listed MAI-Transcribe-1 at $6 per 1,000 minutes on Microsoft Foundry, which is $0.36 per hour. The streaming model is 1.5x that rate, for a different job, and the post-promotion rate is not stated. A cost model built on $0.54 is a floor, not a forecast.

The text-to-speech side: tiered pricing and a Flash variant

Microsoft lists MAI-Voice-2.1 at $22 per million characters and MAI-Voice-2.1-Flash at $15 per million characters, calling Flash about 60 percent cheaper than comparable models, a comparison it does not itemize. The voice models cover 23 languages and 26 locales, and the transcription model is listed at 60 languages with automatic language detection. These are two tier prices on one page, and the vendor's own comparisons are unverified.

What a real test has to measure

A useful pilot replays a few hundred of your own calls and records time to first partial at p50 and p95, how often partials are rewritten, and final word error rate split by accent and noise level, on identical audio against your current provider. The model is listed on Foundry, MAI Playground, Vercel, OpenRouter and Azure Voice Live, with LiveKit coming soon, so the harness needs little new infrastructure.

The bottom line: Wait

Bottom line: Wait. The deciding number is the roughly 100ms to first partial, and it carries no published conditions, while the ranking has no linked method and the price is promotional. Run the replay test now if latency is your bottleneck, and switch only if your own p95 holds up. Watch for independent latency measurements, the post-promotion hourly rate, and a documented AA-WER method.