Microsoft AI's MAI-Transcribe-2 cuts speech-to-text costs 72% at $0.10 per audio hour

Microsoft AI has shipped MAI-Transcribe-2, a speech-to-text model that enters public preview at $0.10 per hour of audio, about 72% below the launch price of the company's previous transcription engine. Released this week, the model records a 5.2% average word error rate across 60 languages on the FLEURS benchmark, a result Microsoft says ranks first in that evaluation.

Microsoft also claims the fastest throughput of any speech model it sells, roughly ten seconds to transcribe an hour of audio, or about 360 times real time. The company puts that at 10x the speed of OpenAI's GPT-Transcribe and 5x the speed of Google's Gemini 3.5 Transcribe. The $0.10 rate is a promotional price valid through December 31, 2026, and Microsoft has not announced what the model will cost after that date.

The accuracy picture is more layered than a single top ranking implies. Microsoft's Azure AI Foundry team notes in its own engineering blog that MAI-Transcribe-2 places second on the Artificial Analysis word error rate leaderboard, even as it contends the model sits on the Pareto frontier of accuracy and latency for the category. Read together, the two benchmarks put the release at or near the top of the market on quality, with throughput the sharper differentiator against rival transcription APIs.

What MAI-Transcribe-2 packs in

The model bundles into a single API capabilities that transcription buyers often had to assemble from separate services:

• speaker diarization that labels each speaker in multi-person recordings
• word-level timestamps for captioning, subtitles, and search
• automatic language identification that needs no language tag supplied up front
• keyword biasing that keeps technical terminology and proper nouns accurate
• code-switching support for mixed-language speech such as Spanglish and Hinglish

Transcripts can be returned in a clean style for readability or a verbatim style when exact wording matters, and Microsoft targets the model at high-stakes documentation work such as clinical note-taking and legal records. Keyword biasing is aimed directly at that audience, letting customers supply lists of drug names, case citations, or product terminology so those strings survive transcription, the point where generic engines typically stumble. Microsoft says accuracy also holds in noisy, real-world recording conditions. For a hospital or law firm evaluating the model, the practical test is whether diarization, biasing, and verbatim output hold up on domain audio without heavy post-editing.

Distribution extends beyond Microsoft's own surfaces. Alongside Microsoft Foundry and the Azure AI Foundry public preview, the model is listed on the MAI Playground and on OpenRouter, so developers outside the Azure ecosystem can reach it through an existing API aggregator rather than a new vendor console. The OpenRouter entry describes a dedicated transcription endpoint with no token context window, a reminder that this is a single-purpose engine rather than a conversational model.

The short history behind the price cut

The discount is steep measured against the line's own history. Microsoft launched the original MAI-Transcribe at $0.36 per audio hour earlier this year, retired it on August 20, 2026, and released a 1.5 refresh covering 43 languages before MAI-Transcribe-2 extended support to 60. That intermediate step had already added content biasing, and the new generation carries the approach further with keyword biasing and code-switching. The promotional rate works out to about $1.67 per 1,000 minutes of audio.

The volume math makes the difference concrete. An operator processing a million hours of audio per year, a plausible figure for a large contact center or media archive, pays about $100,000 at the current rate against $360,000 at the original price. Combined with roughly 360x real-time throughput, that pricing makes full-library transcription economical at volumes that were hard to justify when every hour carried a higher fee.

Ten seconds of compute per audio hour also changes workflow design. A one-hour meeting can be transcribed while it is still running, which makes live captioning and immediate post-meeting notes practical through a single API call, and automatic language identification removes a configuration step for teams handling multilingual recordings. For broadcast and social teams, the word-level timestamps drop straight into subtitle generation pipelines.

The launch is an explicit competitive statement. Microsoft claims MAI-Transcribe-2 undercuts OpenAI, Google, and ElevenLabs on price and speed, and for procurement teams the release resets the reference rate in any speech-to-text negotiation, whatever vendor they currently use. The comparison is pointed because OpenAI's GPT-Transcribe and this in-house Microsoft model now compete for the same developer budgets.

The pace of the product cycle cuts both ways. MAI-Transcribe-1 shipped earlier this year and was retired in August, so early adopters now face a second migration before the year ends. The August 20 retirement landed roughly two weeks before the successor shipped, a handover window short enough to force production users of version 1 to move quickly. Each generation brings re-testing of accuracy on domain-specific audio and re-validation of transcripts in regulated workflows, work that never shows up in the per-hour price.

For global operations the language coverage matters as much as the rate. A single model spanning 60 languages at one price replaces per-region transcription services, and code-switching support handles the alternating-language patterns common in contact centers serving Hinglish speakers and in media interviews that move between Spanish and English.

MAI-Transcribe-2 is part of a wider pattern at Microsoft, which has been building an in-house MAI model family across text, image, voice, and speech rather than licensing models from outside labs. Earlier additions in the same Foundry push include image generation models with a faster MAI-Image-2.5 Flash tier and MAI-Voice-2, a multilingual text-to-speech model with voice cloning across more than 15 languages, with a Flash variant planned. Speech is the corner of that family where the in-house strategy now shows its sharpest price advantage.

For buyers the immediate question is budgeting. The current rate is guaranteed only through the end of 2026, so volume projections should include a scenario with a higher post-promotional price. Teams that move during the promotional window lock in several months of transcription at $0.10 per hour. Given the deprecation pattern, integration work should also pin the API version explicitly and budget for a migration cycle of roughly six months per model generation.

A focused evaluation plan fits inside the promotional window. Buyers can run the model against a held-out sample of their own audio, since leaderboard error rates average over languages and domains, then compare clean and verbatim output with their current vendor on the same files. The same test should confirm that keyword biasing holds the vocabulary that matters for the business, from drug names to case citations to product terminology.

Why this matters

Speech-to-text is being priced as commodity infrastructure, and Microsoft has set the new reference rate for the category. The durable takeaway for enterprises is to capture the savings while the $0.10 window is open and to design architectures that can ride a model line that replaces itself every few months.

Sources

MAI-Transcribe-2: Highest quality transcription, at the fastest speed and lowest cost

New MAI models in Microsoft Foundry across text, image, voice, and speech | Microsoft Community Hub

MAI-Transcribe-2 - Speech Service - Foundry Tools | Microsoft Learn

Related Articles

• Microsoft Launches MAI-Image-2-Efficient for lower AI Costs
• Microsoft MAI Models Cut Copilot Costs in Excel, Outlook
• Microsoft 365 Copilot Gains GPT-5.5 Instant for Faster AI Responses

✔Human Verified

Researched and cross-referenced against primary sources by the Bytevyte editorial team. This article was generated with the assistance of artificial intelligence and reviewed by the Bytevyte editorial team.