Microsoft's MAI-Transcribe-2: a dime per audio hour, and 10x faster than OpenAI
Microsoft is OpenAI's biggest backer, and it just built its own model with the explicit goal of making an OpenAI product look slow.
Three key facts
A 5.2% average word error rate across 60 languages. On September 3, the Microsoft AI team released MAI-Transcribe-2, an in-house speech-to-text model. Microsoft says it ranks first on the multilingual FLEURS benchmark, beating Whisper-Large-V3, OpenAI's GPT-Transcribe, ElevenLabs' Scribe v2 and Google's Gemini 3.5 Transcribe, and places second on the Artificial Analysis word-error-rate leaderboard.
Claimed to be 10x faster than OpenAI. Microsoft says the model runs up to 10 times faster than GPT-Transcribe, 7 times faster than Scribe v2 and 5 times faster than Gemini 3.5 Transcribe. It supports speaker diarization, word-level timestamps, keyword biasing for jargon and proper names, and code-switching such as Hinglish and Spanglish.
$0.10 per hour of audio, through the end of the year. The model is available in Microsoft Foundry, the MAI Playground and on OpenRouter, pitched at captions, meeting notes, call analytics and voice agents.
WangDou's Take
What does a dime an hour buy? A company that sits in meetings eight hours a day would pay under $300 a year to transcribe all of it. That is less than one team offsite's coffee order. Microsoft just turned transcription into tap water.
The price is not the juiciest part, though. The competitor list is. The very first rival named in Microsoft's launch post is OpenAI's GPT-Transcribe, and the chosen metric is the one that stings most: 10x faster. A shareholder with tens of billions sunk into OpenAI is shipping its own models, listing them on a third-party router, and chasing the same developers as its portfolio company. That is not a strategic partnership. That is paying your tenant's rent while building a competing house next door.
Keep some skepticism handy. Vendors pick their own exams, and the promotional price expires at year-end with no word on what comes after. For developers, the smart move is to eat this free lunch for a few months without welding the whole product to one kitchen's menu.
Source: Microsoft AI · Microsoft Tech Community · Reworked