Fahad Manzur
Microsoft AI released MAI-Transcribe-2 on September 3, 2026, a dedicated speech-to-text model priced at $0.10 per hour of audio and available through Microsoft Foundry, the MAI Playground and OpenRouter. Microsoft says the model ranks first on the FLEURS benchmark across 60 languages with an average word error rate of 5.2%, and independent testing by Artificial Analysis clocked it at roughly 10 times the speed of OpenAI’s GPT-Transcribe. The launch price is a limited-time offer that runs until the end of 2026.
What Microsoft shipped
According to Microsoft AI’s announcement, MAI-Transcribe-2 adds speaker diarization, word-level timestamps, keyword biasing for domain-specific terms, and two output styles: a “verbatim” mode that keeps filler words and false starts, and a “clean” mode that strips them out for readable notes and captions. It detects the spoken language automatically and handles code switching, the habit of moving between two languages mid-sentence that Microsoft illustrates with Hinglish and Spanglish. Anyone who has transcribed an Urdu-English business call will recognise the problem.
Microsoft claims the model is up to 10 times faster than leading competitors on long-form audio. In Artificial Analysis evaluations cited in the announcement, it ran 10x faster than GPT-Transcribe, 7x faster than ElevenLabs’ Scribe v2 and 5x faster than Gemini 3.5 Transcribe, while scoring higher on accuracy. The model accepts WAV, MP3 and FLAC files up to 300 MB or two hours long.
The price cut is the real story
eWeek’s analysis puts the new rate in context: $0.10 per hour is about 72% below the $0.36 per hour Microsoft listed for MAI-Transcribe-1 and 1.5. Using Artificial Analysis pricing data, eWeek lists ElevenLabs Scribe v2 at about $0.22 per hour, OpenAI’s GPT-Transcribe at about $0.27 and Google’s Gemini 3.5 Transcribe at about $0.30. On the same leaderboard, MAI-Transcribe-2 recorded a 2.0% word error rate in non-streaming tests, against 2.2% for Scribe v2, 2.6% for Gemini 3.5 Transcribe and 3.3% for GPT-Transcribe.
Two caveats come with that headline number. Microsoft Learn currently lists the model as a public preview with no service-level agreement and does not recommend it for production workloads. And the 10-cent price expires at the end of 2026, with Microsoft yet to say what it will cost afterwards. Anyone building a business around today’s rate should assume it goes up.
Where this lands for transcription freelancers
Transcription and captioning have long been an entry point into online earning for workers in Pakistan, India, the Philippines and Kenya, precisely because the work needed little more than good listening, decent English and patience. That labour is now competing with a model that processes an hour of audio for ten cents, in 60 languages, and labels who said what.
The comparison is not quite as bleak as it sounds. eWeek’s per-language breakdown shows accuracy still varies widely: MAI-Transcribe-2 scores 4.5% word error rate in Chinese and 2.8% in French, but 12.6% in Armenian and 13.1% in Afrikaans, where rival models did better. Urdu, Punjabi, Pashto and Sindhi are not among the languages Microsoft highlighted, and a 5% average error rate still means roughly one wrong word in every twenty. Legal, medical and financial clients cannot ship that unreviewed.
The practical shift is in what clients will pay for. Raw typing of clear English audio is now a commodity that costs less than a cup of tea per hour. What retains value is the layer above it: correcting machine output in noisy or accented recordings, verifying names and technical terms, formatting for court or clinical use, and translating rather than merely transcribing. Freelancers who already use Whisper or similar tools as a first pass, then bill for review and formatting, are positioned for this. Those still quoting per-audio-minute rates for straight typing are competing with a price they cannot match. We covered how AI use should be declared to clients in our post on Fiverr and Upwork’s AI disclosure rules, which becomes relevant the moment a tool like this enters your workflow.
For small businesses in Pakistan and the Gulf, the same release cuts the other way. Meeting notes, customer call summaries, podcast captions and lecture transcripts that were too expensive to produce at scale are now cheap enough to automate, if the language mix is one the model handles well. Testing it on a few real recordings through the free Playground before committing is the sensible first step.
Where does the price go in January?
Microsoft has framed the 10-cent rate as a launch offer, and it has released this model alongside MAI-Code-1.1-Flash and MAI-Image-2.6 in a visible push to sell its own models rather than only reselling OpenAI’s. Whether the price holds, or whether competitors match it before the year ends, will determine if this is a permanent reset of transcription costs or a temporary land grab. Either way, the floor for what a client expects to pay for a plain transcript has moved, and it is unlikely to move back.
