MAI-Transcribe-2-Streaming at $0.54 an hour vs Sume STT per hour

Microsoft's page for MAI-Transcribe-2-Streaming says the model is available at an introductory price of $0.54 per hour of audio through the end of the year. Sume's transcription route, POST /v1/stt-1.0/transcribe, is priced at $0.01 per audio minute, which is $0.60 an hour. On price alone Microsoft's introductory rate is $0.06 an hour lower, or 10 percent, for as long as the offer runs.

Price is not the whole comparison, because the two products are different shapes. Microsoft's is a streaming model for live audio. Sume's is an asynchronous job over an audio file that returns text and word timings, with a 10-minute cap per request. The sections below give the arithmetic from 1 to 1,000 hours and say when each fits.

What did Microsoft announce, and what is reported?

From Microsoft AI's announcement (read 2026-10-03): MAI-Transcribe-2-Streaming at $0.54 per hour of audio, introductory, through the end of the year. The page lists Microsoft Foundry, the MAI Playground and Vercel as places to use the models, with LiveKit marked as coming soon. News coverage on October 1 and 2 repeated the $0.54 figure; the announcement itself is the source used here.

The word introductory matters for budgeting. the sentence read on that page does not say what the rate becomes after the year ends, so a spreadsheet that extends $0.54 into 2027 is guessing.

What does Sume charge per hour of transcription?

Sume's catalog entry for sume/stt-1.0 is $0.01 per audio minute after the standard 1.25 margin on the provider's $0.008 list. The duration_seconds field on the request is used for the reservation; omit it and Sume reserves for one minute. The maximum is 10 minutes, so the largest single reservation is $0.10 and an hour of audio is six requests of up to 600 seconds each.

There is no live-stream mode. The route returns a job (mode: async by default, with sync and subscribe waiting at most 30 seconds), and the result carries text plus words[] with start and end seconds. For a long recording you cut the audio into pieces of at most 10 minutes first; Sume's Timeline audio job can split one hosted file into ranges reserved at $0.01 per job, and you submit each range to STT.

What does an hour, a day and a month of audio cost?

Straight multiplication of each published rate. Sume's figures ignore the split step, which adds one flat $0.01 job per source file if you use Timeline audio.

When does the difference not matter?

A $60 gap on 1,000 hours is smaller than the engineering time to rebuild a pipeline, so the decision usually turns on shape rather than price. If the product needs captions while someone is still speaking, Sume's async job does not do that, and a streaming model is the right tool. If the audio already exists as files, such as call recordings, interviews or video soundtracks, an async job with word timings is a better fit, and it sits next to Sume's other steps: the same API key and USD balance also pay for captions, trims and renders.

One thing the Sume side makes easy is spend control, because the reservation is known before the job starts. A 10-minute file reserves $0.10 and a 90-second clip reserves $0.015. A call fails with 402 insufficient_credits before any transcription starts if the balance cannot cover it.

• Live captions, voice agents: needs streaming; Sume STT is not streaming.
• Batch transcripts with word timings: Sume STT, 10 minutes per request.
• Budget past December 31, 2026: the page read does not state the post-introductory rate.
• Language handling: Sume takes an optional language_code hint and auto-detects when omitted.

How do I submit a Sume transcription and cap the spend?

Set duration_seconds to the true length of the clip, or the reservation defaults to one minute and your balance check will be wrong for longer files. An Idempotency-Key makes a retry safe.

A final check before you compare any two speech prices: confirm what each figure covers. The vendor figure is an introductory per-hour rate for a streaming product, while the Sume figure is a per-minute rate for a file you submit, so the two numbers describe different ways of using speech to text. Compare them on your own audio hours, not on the headline numbers.

If you transcribe files rather than live audio, the Sume maximum is the number to plan around: a 10 minute file reserves $0.10, so an hour of audio split into six such files reserves $0.60 in total. Pass duration_seconds on every call so the reservation matches the file rather than defaulting to one minute.

curl -X POST https://api.sume.com/v1/stt-1.0/transcribe \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: stt-part-01" \
  -d '{
    "audio_url": "https://example.com/part-01.mp3",
    "duration_seconds": 600,
    "language_code": "en",
    "mode": "async"
  }'