Contents
On September 3, Microsoft AI shipped MAI-Transcribe-2, a speech-to-text model that ranks first on the FLEURS multilingual benchmark and second on the Artificial Analysis Word Error Rate leaderboard, while pricing itself at $0.10 per audio hour through the end of 2026. That is not a small cut. The previous generation launched at $0.36 per hour, so the sticker price for transcription just dropped by roughly 72% while the model got more accurate, not less.
This matters because transcription has quietly become a load-bearing piece of a lot of AI pipelines: meeting summarizers, voice agents, captioning tools, and compliance recordkeeping all sit on top of a speech-to-text layer that most teams treat as a commodity utility rather than something worth benchmarking. When the cost of that utility drops by two-thirds and the accuracy goes up at the same time, it changes which workloads are worth running at all, especially the ones that were previously too expensive to transcribe at scale.
What actually changed
Pricing dropped from $0.36 to $0.10 per audio hour. That works out to about $1.67 per 1,000 minutes, and Microsoft has committed to holding that rate through December 31, 2026, after which it says a non-promotional price will follow. There is no separate charge for the new structural features, diarization and timestamps ride along in the same per-hour rate.
Accuracy and speed stopped trading against each other. MAI-Transcribe-2 sits at No. 1 on Artificial Analysis’s accuracy times latency Pareto frontier, meaning no other model available today beats it on speed at the same accuracy or on accuracy at the same speed. In Microsoft’s own non-streaming tests it posted a 2.0% word error rate at a median speed factor of 410.7x real time, and the company claims it is roughly 10x faster than OpenAI’s GPT-Transcribe, 7x faster than ElevenLabs Scribe v2, and 5x faster than Google’s Gemini 3 on comparable transcription jobs.
Language coverage and structural features both expanded. Support grew to 60 languages with automatic language identification and code switching for mixed-language speech, and the model adds speaker diarization and word-level timestamps, the two capabilities Microsoft says production teams asked for most after the first generation shipped without them. It is live today in public preview through Azure AI Foundry and Azure Speech.
Try it: what the price cut is worth on your workload
About the calculator
Uses Microsoft’s published $0.10/hour promotional rate for MAI-Transcribe-2 against the $0.36/hour launch price of MAI-Transcribe-1. Both are list prices for the batch transcription API; your actual bill depends on which tier and features you use.Move the slider to see the monthly cost difference.
At 5,000 hours a month, a mid-size support or meeting-transcription pipeline, the switch alone saves about $1,300 a month at list price, before counting the added value of diarization and timestamps that used to require a second API call.
What this means for you
If you are running a speech-to-text pipeline on any of the incumbents, the honest move is to benchmark MAI-Transcribe-2 against your own audio before switching on price alone, word error rate on FLEURS is a strong signal but it is not your call center’s accent mix or your product’s background noise. The diarization and timestamp support closing the gap with dedicated tools like ElevenLabs is the more durable change here, since it removes a whole second vendor from pipelines that used to need one model for text and another for speaker attribution. My own next step is running a side-by-side transcription test on our own support-call sample before deciding whether to route any of our audio pipelines through Azure AI Foundry.