7 hours 13 minutes of audio: transcript bill on MAI vs Sume STT

Transcribing 7 hours 13 minutes (433 minutes, 7.2167 hours) of audio costs about $0.72 on MAI-Transcribe-2 batch, $3.90 on MAI-Transcribe-2-Streaming and $4.33 on Sume STT, using Microsoft's introductory page prices and Sume's $0.01 per audio minute. On Sume it takes 44 requests, because one STT request is capped at 600 seconds.

The bill by model

The rates are the Microsoft model page's per-hour prices and the Sume catalog's per-minute price, read 2026-10-08. Sume's total is 433 x $0.01 = $4.33. The Microsoft totals are 7.2167 hours x the hourly rate.

Why Sume needs 44 requests

Sume STT's estimate and request cap is 600 seconds. 433 minutes divided by 10 is 43.3, so you need 43 requests of 10 minutes and one of 3 minutes: 43 x 10 + 3 = 433. If the source is a video, audio detach can cut a range out of the video's audio at $0.01 per job; its documented shape for STT is 16 kHz mono, and one detached range may not exceed 900 seconds, so ranges of 600 seconds fit both limits. Cutting 433 minutes of video audio into 44 ranges would add 44 x $0.01 = $0.44 of detach jobs, taking the Sume total for video sources to $4.77. If your files are already audio, there is no detach step.

Each request carries a duration_seconds field you can set; the catalog says that omitting it reserves one minute, so set it per chunk so the reservation matches the audio.

Reading the result

The cheapest row is not automatically the right one. The batch model is $0.10 per hour on the page's introductory price and the page does not say what the price becomes afterwards. The streaming model is built for live audio and, on the page's feature rows, returns no word timestamps or diarization. If you are transcribing recordings after the fact, streaming is the wrong tool at $3.90; if you want the transcript inside Sume's job and Timeline flow, the $4.33 includes that integration and not just the text.

A worked per-chunk plan

For a folder of recordings you can plan the Sume side as a list. Suppose the 433 minutes are 13 files: twelve files of 33 minutes (396 minutes) and one of 37 minutes. Each 33-minute file needs 4 requests (10 + 10 + 10 + 3) and the 37-minute file needs 4 (10 + 10 + 10 + 7), so the job count is 12 x 4 + 4 = 52, not 44, because chunking per file cannot merge the short remainders. The audio minutes, and so the $4.33, do not change; only the number of requests and any per-request overhead do.

That is the useful planning rule: the Sume price depends on total minutes, while the request count depends on how the minutes are spread over files. If you control the source, concatenating recordings with Timeline audio (up to 20 parts per job, $0.01) before splitting at 600 seconds can bring the count back toward 44, at a cost of the extra join jobs.