MS Releases MAI-Transcribe-2… Speech Recognition Model 72% Cheaper and 5–10x Faster Than Competitors

Microsoft has launched the speech recognition model MAI-Transcribe-2 as part of its in-house frontier model strategy. Priced at $0.10 per hour, the model achieves approximately 72% cost savings compared to the previous version, reducing costs from $36,000 to $10,000 for processing 100,000 hours annually.

Performance and Feature Highlights

MAI-Transcribe-2 supports 60 languages and natively provides code switching capabilities, allowing language changes mid-sentence in formats like Spanglish and Hinglish. Additionally, features previously sold separately at premium tiers by competitors—such as speaker diarization, word-level timestamps, and keyword biasing—are included in the base pricing.

In terms of performance, it ranked first on the FLEURS benchmark with an average error rate of 5.2%. In Artificial Analysis evaluations, it demonstrated processing speeds 10x faster than OpenAI's GPT-Transcribe, 7x faster than ElevenLabs' Scribe v2, and 5x faster than Google's Gemini 3.5 Transcribe.

Strategic Intent and Market Impact

Microsoft has been developing its own models to reduce reliance on OpenAI and secure margins. Following the termination of exclusive access to OpenAI models via contract revisions in April 2026, Microsoft began integrating MAI models into its own products such as Word and Excel. The speech transcription market was selected as the first target for in-house model replacement due to its limited problem scope and objective metrics.

A focused team of 10 led by Mustafa Suleyman reduced GPU costs by half compared to competitors through a development structure that eliminated bureaucracy. This is interpreted as an intent to commoditize transcription into a low-price (high-volume) offering and capture the market before competitors can respond.

Remaining Challenges and Limitations

The launch announcement leaves several practical gaps. It is unclear whether the $0.10 per hour price is a launch offer, as no end date or standard rate is specified. Furthermore, there is no mention of real-time streaming transcription, and detailed metrics and policies required by enterprise customers—such as diarization error rate (DER), data residency, and training data usage—have not been disclosed. Amidst a trend where some US companies are evaluating cheaper Chinese models despite security concerns, Microsoft aims to gain a competitive edge by emphasizing compliance and reliability.