Transcription is one of those AI markets that rarely makes headlines but quietly underpins a lot of others: meeting notes, media captioning, accessibility tools, and the voice agents everyone is now racing to build all sit on top of a speech-to-text model. This week Microsoft dropped a new one into that market and priced it to hurt.

MAI-Transcribe-2, released on 3 September by Microsoft AI — the company's in-house model division — is being pitched, in Microsoft's own words, as "the fastest, most accurate and cheapest speech recognition model in the world." The claims behind that line are unusually specific, which makes them easy to check as independent testing catches up.

The numbers

On accuracy, Microsoft says the model ranks first on the FLEURS benchmark across 60 languages, with an average word error rate of 5.2%. FLEURS is a standard multilingual speech test, so a top ranking there is a real claim rather than a vague one — though, as ever, the ranking is Microsoft's own until third parties reproduce it.

On speed, the pitch is aggressive: 10× faster than OpenAI's GPT-Transcribe, 7× faster than ElevenLabs' Scribe v2, and 5× faster than Google's Gemini 3.5 Transcribe. Microsoft says it also beats OpenAI's older Whisper V3-Large and ElevenLabs' ScribeV2 across a broader spread of real-world audio.

And on price — the part that reorders markets — it lists at $0.10 per hour of audio, described as a limited-time rate until the end of the year. Transcription APIs from the incumbents have generally cost several times that. Undercutting the field by that margin on a model that also claims the top accuracy score is a land-grab, not a me-too launch.

The feature set is built for real work rather than demos: speaker diarization, word-level timestamps, keyword biasing, verbatim and clean transcription modes, code-switching between languages, automatic language identification, and robustness in noisy conditions. It is available now through Microsoft Foundry, the MAI Playground, and OpenRouter.

The part that is really the story

The interesting thing is not the transcription model. It is whose model it is. MAI is Microsoft's own model effort — the group building the company's in-house systems rather than reselling its partner's. Microsoft remains OpenAI's largest shareholder, and OpenAI's models still run through Azure and Copilot. But a Microsoft model that is explicitly benchmarked as beating GPT-Transcribe is Microsoft competing, in public, with the company it is most tied to.

That is the pattern worth watching. Speech-to-text is a contained, measurable slice of the market — a good place to prove an in-house model can top the incumbents on quality and price at once. If it holds up under independent testing, it is a template Microsoft can run again in other categories where it would rather own the model than rent it.

Read it with the usual caveats

Two are worth stating. The benchmark and speed numbers are, for now, Microsoft's own — and even on Microsoft's own page the accuracy claim comes with an asterisk: the model ranks first on FLEURS but only second on the independent Artificial Analysis word-error-rate leaderboard. "Most accurate in the world" is a headline that already needs a footnote. And a "limited-time" $0.10 price is a marketing lever as much as a cost: the number that matters is what it settles at once the launch push is over. Neither undercuts the basic point — a major lab just shipped a cheap, fast transcription model and aimed it squarely at three rivals at once — but both are reasons to watch the independent numbers rather than the press release.