Transcription drops to 10 cents an hour

Microsoft AI released MAI-Transcribe-2 on Thursday and set the price at $0.10 per hour of audio, roughly 72% below the $0.36 an hour it charged for the first model in the line five months ago. The model is available through Microsoft Foundry and MAI Playground, VentureBeat reported.

At a glance

• Speaker diarization, word-level timestamps, keyword biasing, automatic language identification and mid-sentence code switching between languages such as Hinglish and Spanglish ship inside the base rate rather than as premium add-ons.
• Coverage runs to 60 languages, up from 43 in June and 25 in April, and Microsoft claims first place on the FLEURS multilingual benchmark with an average word error rate of 5.2%.
• For an enterprise processing 100,000 hours of call-center audio a year, the annual transcription bill falls from $36,000 at the old rate to $10,000 at the new one.

The pricing reads as a substitution play as much as a product launch. Every hour of audio that Microsoft moves from a partner's model to its own removes a payment it currently makes, and the surplus capacity is what it is now selling at a dime. Selling the same capability at that level, with diarization and timestamps included, likely squeezes the specialist transcription vendors that charge separately for both.

MAI-Transcribe-2 covers 60 languages and bundles diarization and timestamps in the base rate

The model transcribes audio in 60 languages, against 43 in June's MAI-Transcribe-1.5 and 25 in the original April release, and runs on Microsoft Foundry, the company's model marketplace for developers, and in MAI Playground, its testing environment. Microsoft says it built the model for noisy recordings, low-quality audio and overlapping speech rather than studio conditions.

Speaker diarization attributes speech to individual participants, word-level timestamps mark each word for search and video alignment, and keyword biasing lets developers supply lists of drug names, product codes or employee names. A configurable output style offers a verbatim mode that keeps fillers and false starts for compliance and legal work and a clean mode that strips them; code switching covers conversations that shift language mid-sentence, including Hinglish and Spanglish.

Microsoft claims first place on FLEURS at a 5.2% word error rate

Microsoft says MAI-Transcribe-2 ranks first on FLEURS across the 60 languages with an average word error rate of 5.2%. FLEURS, published by Google researchers in 2022, is built from native speakers reading roughly 2,000 sentences in each of 102 languages, about 12 hours of read speech per language.

That 5.2% figure is above the 3.7% Microsoft reported for MAI-Transcribe-1.5 in June, which was averaged across 43 languages rather than 60. On the Artificial Analysis word-error-rate leaderboard Microsoft claims second place and the accuracy-latency Pareto frontier; in June that firm ranked MAI-Transcribe-1.5 third at 2.4% WER, behind Alibaba's Fun-Realtime-ASR-preview and ElevenLabs' Scribe v2.

Per Artificial Analysis evaluations, the model runs 10 times faster than OpenAI's GPT-Transcribe, seven times faster than Scribe v2 and five times faster than Google's Gemini 3.5 Transcribe. Microsoft's release names those systems plus OpenAI's older Whisper V3-Large, and does not mention Deepgram, AssemblyAI, Speechmatics or Rev.

A 10-person team built the line as the OpenAI partnership loosened

MAI-Transcribe-1 launched on April 2 with 25 languages at $0.36 per hour, and MAI-Transcribe-1.5 followed on June 2 with 43 languages. Mustafa Suleyman, chief executive of Microsoft AI, told The Verge in April that a "small, focused 10-person team" built the first model at "half the GPU cost of the other state-of-the-art models".

Microsoft has invested more than $13 billion in OpenAI. Its October 2025 restructuring of the partnership allowed Microsoft to "independently pursue AGI alone or in partnership with third parties", per the company's own announcement, and reports at the time said an April 2026 amendment ended Microsoft's exclusive access to OpenAI models and its revenue-share payments.

Bloomberg reported in July that Microsoft had begun using MAI models to answer a portion of user prompts in Word and Excel. Microsoft also owns Teams, Nuance and the Azure speech services, workloads that generate or process large volumes of enterprise audio.

Streaming, and when $0.10 ends Microsoft describes $0.10 per hour as a launch offer and names neither an end date nor a standard rate. The release emphasizes batch throughput and long-form audio and says nothing about real-time streaming, where Meta's Muse Voice Transcribe currently leads. It also gives no per-language error breakdown, no diarization accuracy metric and no detail on data residency, retention or training use.