MAI-Transcribe-2 drops transcription to $0.10 per hour

Microsoft AI shipped MAI-Transcribe-2 on September 3 at $0.10 per hour of audio. That is not a typo, and it is not a rounding error against the previous generation: MAI-Transcribe-1 launched in April at $0.36 per hour (Microsoft AI). Speech-to-text just stopped being a budget line for most small operators. What it did not stop being is a dependency.

What actually happened

MAI-Transcribe-2 covers 60 languages with automatic language identification and posts a 5.2% average word error rate on FLEURS, which Microsoft says puts it first on that benchmark and second on the Artificial Analysis WER leaderboard. Microsoft claims it runs 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs' Scribe v2, and 5x faster than Gemini 3.5 Transcribe — vendor benchmarks, so treat the multiples as directional.

The feature list is the part that matters operationally: speaker diarization, word-level timestamps, keyword biasing for domain vocabulary, code switching for mixed-language speech, and configurable verbatim or clean output. It is live through Microsoft Foundry, the MAI Playground, and OpenRouter.

One caveat worth reading twice. Microsoft frames the $0.10 rate as a limited-time price through the end of 2026, and no standard rate is published anywhere — VentureBeat flagged the same gap. You are budgeting against a number with an expiry date and no successor.

Why cheap transcription changes your automation math

At $0.10 an hour, workloads you shelved on cost come back. Every sales call, every support line, every site walkthrough, transcribed and searchable, for less than the coffee. A thousand hours of call audio a month is a hundred dollars. The question stops being "can we afford to transcribe this" and becomes "what do we do with the text."

Which is exactly where teams get stuck to a vendor. The model is interchangeable — WER differences at the top of the leaderboard are inside the noise of your recording quality. The lock-in is everything downstream: the shape of the diarization payload, the timestamp format your search index expects, where PII redaction happens, which fields your CRM writes back. Wire those to one provider's response schema and swapping costs you a sprint instead of a config line.

Put an adapter between the API and your pipeline. Normalize to your own transcript object — speaker, start, end, text, confidence — and let the provider be a swap. Then run the same 20 recordings through two vendors and score them on your audio, not on FLEURS. Accented speech, crosstalk, a noisy shop floor, and your product names are what actually decide accuracy.

Sources: Microsoft AI, VentureBeat.