MAI-Transcribe-2-Streaming intro rate runs to year end: budget plan

Microsoft's page for MAI-Transcribe-2-Streaming states the price as "$0.54 per hour of audio through the end of the year", an introductory rate (read 2026-10-07). The page does not say what the rate becomes afterward, so any budget that runs past December should treat $0.54 as a number with an expiry date, not a long-term line item.

Sume STT 1.0 is $0.01 per audio minute, which is $0.60 per hour, listed in STT_PUBLIC_PRICING and readable live from GET /v1/catalog. The two prices are close, and the interesting part is how each behaves over a calendar.

The arithmetic at four volumes

Both prices scale linearly with audio hours, so the monthly difference at the intro rate is simply hours multiplied by $0.06. At the intro rate Microsoft is cheaper per hour; the gap is small enough that it vanishes if you value file-based jobs, a single API key, or word timings you can feed straight into captions.

What changes on January 1

Nothing in the Microsoft page says the price rises, falls or stays put after the intro period. The only safe plan is to model both directions. If the rate went up by 20 percent, $0.54 would become $0.648 per hour, which is above $0.60; if it stayed, you keep the $0.06 advantage. That is a thought experiment, not a forecast, and you should check the vendor page again in December. Put the review date in your calendar now, because a budget approved in October and reviewed in March is the usual way an intro price turns into a surprise.

The Sume side has no expiry in its rate card, but it is not frozen either: the catalog is the source of truth, and a budget script should read the rate at run time instead of hardcoding it.

Streaming and batch are not the same purchase

The streaming model exists to deliver partial transcripts as audio arrives, which matters for voice agents. A pile of recorded clips does not need that, and a recorded clip is cheaper to think about as a batch job: submit audio, poll, read text and word times. If your workload is subtitles for finished videos, price it as batch work first and ask whether the streaming latency buys you anything.

Sume's file-based path for a video is detach the audio, transcribe, then burn. Audio detach is $0.01 per job and outputs a 16 kHz mono WAV in the STT shape, and the transcription job is capped at 10 minutes of audio per request, so a long file is several jobs.

Questions to ask before you commit

Is the work live or recorded? If it is recorded, the streaming premium in engineering effort buys nothing. Is the audio under 10 minutes per file? Sume's job limit is 600 seconds, so longer files need chunking, and that adds a small fee per slice. Do you need captions as well? The transcript is only half of a subtitle: the burned video is a separate $0.20 job on Sume for up to 60 seconds, and a vendor's per-hour STT rate does not include rendering.

Where will the text go? Microsoft lists Foundry, the MAI Playground and Vercel as places to use the streaming model, with LiveKit listed as coming soon (read 2026-10-07). If your pipeline already lives in one of those, switching cost matters more than a six cent hourly gap.

A budget script that survives a price change

Keep the rate in one place. Read it from the catalog before each batch, multiply by audio minutes, and add the detach fee per source file. For a 30 minute video, the audio is three 10 minute jobs. STT is 30 x $0.01 = $0.30, and three detach ranges are 3 x $0.01 = $0.03, so the total is $0.33 before captions. Microsoft's intro rate for the same half hour would be 0.5 x $0.54 = $0.27 of transcription, a gap of $0.03 on STT alone. The gap is real but small, and the decision usually turns on integration cost, not cents.