MAI-Transcribe-2: speech-to-text for $0.10 an hour, top of FLEURS

Microsoft AI released MAI-Transcribe-2 in early September, and the number that matters is on the price list: $0.10 per hour of audio, valid through the end of 2026, down 72% from $0.36. It covers 60 languages, does speaker diarization, and returns word-level timestamps. It also sits at number one on FLEURS with 5.2% WER, and Artificial Analysis measured it at roughly 10x faster than GPT-Transcribe, 7x faster than ElevenLabs Scribe v2, and 5x faster than Gemini 3.5 Transcribe. Cheaper, faster, and more accurate at once is the combination that turns a product category into a utility. Speech-to-text is now a utility.

What shipped

• Price: $0.10 per audio hour through end of 2026, from $0.36 before.
• Coverage: 60 languages, diarization, word-level timestamps.
• Accuracy: first on FLEURS at 5.2% WER.
• Speed: 10x, 7x, and 5x faster than GPT-Transcribe, Scribe v2, and Gemini 3.5 Transcribe respectively, per Artificial Analysis.
• Distribution: Microsoft Foundry, the MAI Playground, and OpenRouter.

The OpenRouter route is the one I would test first: same request shape as everything else you route, no separate vendor onboarding. The unified API covers how that works if you have not used it for audio before.

Do the arithmetic once

Take a call-centre client with 100,000 hours of audio a year. At $0.36 that was $36,000. At $0.10 it is $10,000. That $26,000 is not a rounding error on the transcription budget; it is the whole margin most middleware vendors were living in. If you resell transcription, your invoice just got audited by a press release. If you buy it, rebid this quarter.

Speed matters as much as price for batch and streaming pipelines. A 10x real-time factor improvement means the same worker pool clears ten times the backlog, or you run a tenth of the workers. Diarization and word timestamps in the base product remove two post-processing steps that used to justify a separate service.

When the best model on the leaderboard is also the cheapest one on the price list, the transcription layer stops being a place you can add value and starts being a place you can only add cost.

Where the value moves

If the STT layer is a commodity, the work is everywhere around it.

• Workflow. Getting audio in, getting text to the right system, handling retries and partial failures, and doing it without a human copying files. Nobody pays for a transcript; they pay for what happens next.
• Compliance. Where the audio goes, where the transcript is stored, how long, and who can read it. For EU clients that is the first question, not the last, and a first-party Microsoft endpoint through Foundry may be easier to clear than a startup's API.
• Domain data. Product names, drug names, internal jargon, and the correction loop that teaches your downstream systems what the model got wrong. The model is generic; your vocabulary is not.
• Routing. Not every minute needs the top model. If you already tier LLM traffic, cost routing applies to audio the same way: cheap model for the bulk, best model for the calls that matter.

What to verify before you switch

• Czech quality. FLEURS covers many languages, but a headline WER is an average. Run your own 50-file sample in Czech, with your speakers, and measure WER and diarization error yourself.
• Diarization on real calls. Cross-talk, hold music, and two people with similar voices break diarizers that look perfect on clean benchmarks.
• Word timestamp precision if you drive captions or redaction from them.
• What happens after 2026. The $0.10 rate is time-boxed. Build the cost model with a reversion to the list price and see if the switch still pays.

The honest gap

Every performance figure is vendor-published or from Artificial Analysis, and the price is a promotional rate with an expiry date. I have not run MAI-Transcribe-2 on Czech audio, and FLEURS leadership tells you nothing about noisy multi-speaker telephony. The economics are real enough to act on; the quality claims are the part you test.