Reference ingest cost: free manifest, STT only if speech is heard

Sume's reference ingest costs nothing for the manifest, which covers shots, text read from frames and audio facts for a clip of up to 300 seconds. The one billed part is an optional transcript, reserved at $0.01 per audio minute and settled to zero if the track is silent or has no speech.

The surface is dest first: the docs say it is listed where its flag allows it, with development auto-on and production opt-in. Check the tool list in your environment before you plan around it.

What is billed and what is not?

The reference ingest docs say the manifest is unbilled, CPU work on the media runtime like video frames. Transcription is the exception, and it is off unless you opt in with speech.allow_billed_stt.

A clip over 300 seconds is refused with source_too_long_for_reference_ingest, and the error points you to video inspect.

How does the transcript reservation settle?

With speech.allow_billed_stt: true, Sume reserves the sume/video-inspect-1.0#transcript rate, $0.01 per audio minute, per ceil(minute) of your duration_seconds hint, or one minute when the hint is absent. The hint maximum on this route is 300 seconds. It transcribes only when the track is not silent and voice activity detection finds speech. Otherwise it settles to zero and reports stt_skipped_silent, stt_skipped_no_speech or stt_skipped_no_audio_track.

So the ceiling for a transcript on a 300-second clip is five minutes, or $0.05, and a music-only clip costs nothing even if you opted in.

What does a batch of references cost?

The arithmetic is simple because only speech minutes bill. These rows use the published $0.01 per minute and assume every clip contains speech, which is the worst case.

Short clips round up to a minute each, which is why 100 half-minute clips cost $1.00. If some are music-only, the real figure is lower.

How do I keep the bill at zero?

Leave speech.allow_billed_stt out. The manifest still carries the shot list, the text tracks and the audio gate, and audio.silent: true tells you to plan new music and not to request STT at all. The docs say speech.language_code and duration_seconds without the opt-in are rejected with reference_ingest_stt_required, so you cannot pay by accident through a stray field.

Idempotency-Key is required on the submit.

What does it not do?

Ingest is a deterministic preflight, not a semantic analysis. semantic: true is refused with reference_ingest_semantic_unavailable until an enrichment pass ships. It reads one clip your workspace already holds on media.sume.com and produces no MP4, so there is nothing to generate and nothing to refund on the unbilled path.

If the surface is not listed for your environment, video inspect is the fallback the docs name, with the same $0.01 per minute transcript rate.

Where does the cost show up in usage?

The docs say the transcript reservation settles to what ran, so a job that skipped transcription settles to zero and the warnings in the manifest tell you why. The docs excerpt we read does not describe how the charge is itemised in the usage ledger, so read your own usage records for the final figure.

For budgeting, assume zero for the manifest and cap the transcript at $0.05 per clip of up to 300 seconds.