Gemini 3.5 Transcribe Live pricing and Sume's async STT job

Google's pricing page lists Gemini 3.5 Transcribe Live at an estimated blended rate of about $0.009 per minute, for bidirectional streaming over WebSockets. Sume's speech transcription is a different shape: you send a finished audio file as a job and it is priced per audio minute, at $0.01 per audio minute on the rate card.

Both are per-minute numbers, but they buy different things, so compare the workflow before the rate. Google figures are from its page and Sume figures from the rate card and Video inspect docs, read 2026-09-30.

How does Google price the Live model?

The page shows per-token prices: $3.50 per 1M input tokens (about $0.005 per minute of audio) and $21.00 per 1M output tokens (about $0.004 per minute of text). It states the estimate uses 25 audio tokens per second for input and 175 text tokens per minute for output, giving about $0.009 per minute.

How is Sume's STT job billed?

Sume STT 1.0 is priced per audio minute after Sume margin. You can send a duration_seconds hint; if you omit it, Sume reserves 1 minute, and the maximum hint is 600 seconds. Paid usage is reserved at submit and captured on completion, as the generation admission docs describe.

Which one fits my use case?

If you need words while a person is still speaking, streaming is the shape Google's Live model is built for. Sume does not document a streaming transcription endpoint, so for live captions use a streaming service. If you have recordings, a job with a duration hint gives you a known reservation before work starts. For long files see transcribe long audio files.

What should I check before comparing prices?

Confirm the unit (tokens versus audio minutes), whether output tokens are billed, and the file length limits. Sume's own limits are covered in Gemini 3.5 Transcribe limits versus Sume STT. Read the live rate with GET /v1/catalog rather than copying a number from a blog.