MAI-Transcribe-2-Streaming vs Sume STT: 100 hours, $54 vs $60

The numbers

For 100 hours of audio, the vendor listing works out to 100 x $0.54 = $54.00 and Sume STT to 100 x 60 minutes x $0.01 = $60.00. The listed MAI rate is introductory through the end of 2026, per the tracker. Sume does not list MAI-Voice-2.1, MAI-Voice-2.1-Flash or MAI-Transcribe-2 in its catalogs, so this is a price comparison, not a statement that you can call Microsoft's models through Sume.

Side by side

Vendor prices come from a third-party release tracker that lists them, not from a Microsoft price page, so confirm them on Microsoft's own pricing before you commit.

What the price does not compare

Accuracy, latency, language coverage and streaming support are not compared here. The tracker lists 60 languages for the Microsoft model; Sume's docs do not state a language count, only that language_code is an optional hint and omission means auto-detect.

Sume STT returns word timings always, and sentence segments if you request segmentation, which drop into caption lines. Provider options such as diarization are fixed server-side, so if you need to tune them this is not the surface.

Which one when

If you only need transcripts at volume, the per-hour gap is $6 per 100 hours and workflow should decide. If transcripts feed Sume captions, clip inspection or a Timeline render, using Sume STT keeps everything in one job log. Sume STT bills $0.01 per audio minute on video inspect with transcribe true (Video inspect); the audio source format is described in Audio detach.

What changes at larger volume

At 1,000 hours the gap grows to $540 versus $600, a $60 difference. The vendor rate is introductory through the end of 2026, so any plan that runs into 2027 should not assume it. Sume's rate is documented as a public rate with the live price in GET /v1/catalog.

Costs outside transcription may dominate: storage, review time and downstream edits are usually larger than a six-dollar gap per hundred hours.

Where Sume's STT is used

Sume uses the same STT inside video inspect, reference ingest and captions, which is why it is priced per audio minute. In reference ingest, allow_billed_stt true reserves the per-minute rate and transcribes only when the track is not silent, and silent tracks settle to zero. That is a guard worth copying in any pipeline that processes many files of unknown content.