Video inspect transcribe: per-minute STT add-on, 10-minute cap
Adding transcribe true to a Sume video inspect adds $0.01 per audio-minute of your duration hint, one minute if omitted, up to 10 minutes.

How is the transcript on a Sume video inspect billed? Only transcribe: true bills, at $0.01 per audio minute of the duration_seconds hint you send. If you omit the hint, Sume reserves one minute, and the hint maximum is 600 seconds, so the transcript add-on tops out at $0.10.
This follows the Sume video inspect docs, which also say the add-on sits on top of the inspect's own compute reservation.
The rule in one line
The catalog entry for the transcript reads: per ceil of the audio minute of the hint, one minute when absent, maximum 10 minutes, frozen at the legacy STT 1.0 public rate. So the add-on is the number of minutes, rounded up, times $0.01. A 61-second hint is two minutes.
| duration_seconds hint | Minutes billed | Add-on |
|---|---|---|
| Omitted | 1 | $0.01 |
| 45 | 1 | $0.01 |
| 61 | 2 | $0.02 |
| 180 | 3 | $0.03 |
| 600 (maximum) | 10 | $0.10 |
What the hint is and is not
The hint sizes the reservation. It is not a cut: the docs describe duration_seconds as a hint for usage, so send the true clip length rounded up rather than a guess. The video inspect page applies the 600 second maximum to the hint, so a longer clip needs a plan, for example detaching the audio and transcribing ranges separately.
The inspect itself reserves its compute ceiling and is captured at its own container seconds, never above the hold. The transcript's per-minute rate is added to that reservation, so budget the two lines separately.
Fields that only work with transcribe
Three fields apply only with transcribe: true: language_code, segmentation and duration_seconds. Sending any of them without it returns 400 video_inspect_transcribe_required, which is a cheap error to hit in testing.
language_code is a hint, and omitting it means auto-detect. segmentation.mode: "sentence" returns sentence segments shaped like caption lines.
Check for audio first
A clip with no audio track fails with inspect_source_has_no_audio. Run an inspect with frames: false first and read probe.has_audio; that is enough to know whether a transcript can succeed before you pay for one.
The route also defaults to sync mode, waits up to 30 seconds and answers 200 with the finished inspect, or 202 with a queued job to poll.
Budgeting a batch
Suppose you inspect 50 clips and send a 90 second hint for each. Each rounds up to two minutes, so the transcript add-on is 50 times $0.02, which is $1.00, plus the compute line for each inspect. If you leave the hint out, the reservation is one minute per clip, so check that a real clip of that length is covered before relying on it.
Send a hint that matches the clip. A hint that is too low under-reserves, and one that is too high holds more than you need for the duration of the job.
Where to confirm the price
Rates are catalog-driven, so confirm the live number with GET /v1/catalog, covered in the API reference, before you budget a large batch.
Sources
Related posts
More in Pricing
- Video Router: what a request with no resolution or duration costs
Leave out resolution and duration and Sume prices a default. Wan 3.0 reserves a 1080p, 5-second clip. See each model's default and cost per clip.
- Video to transcript cost: audio detach plus STT, by minutes
Detach a video's audio for $0.01, transcribe it at $0.01 a minute. One-, three-, five- and ten-minute videos cost $0.02 to $0.11 at Sume list rates.
- Wallet needed to fill the Sume job queue, by plan
Filling Pro's 24 accepted job slots with 30-second Standard avatar clips reserves $132.48 up front; Scale's 120 slots reserve $662.40. Free has 6 slots.
- Wan 3.0 and MiniMax H3 Max clip price on Sume at 5, 10 and 15 seconds
At list times 1.25, a 10-second Wan 3.0 clip is $0.625 at 480p, $1.25 at 720p and $2.50 at 1080p. MiniMax H3 Max is $0.625, $1.00 and $2.00 at its three tiers.
Written by Sume