Check probe.has_audio before paying $0.20 for Short captions

Probe first. A Video inspect call with frames set to false returns probe facts without stills, and the docs say that is enough to read probe.has_audio. If the clip has no audio, speech-to-text captions will fail with caption_no_speech, and you should send authored cues instead. A standalone caption job is $0.20 for clips of up to 60 seconds (as of 2026-10-08), so for a 59-second Short a wrong call costs a fifth of a dollar plus a retry.

Two different failure codes

Video inspect and Video captions guard the same fact from two sides. Inspect with transcribe true on a silent clip returns inspect_source_has_no_audio. Captions without cues or words, on a silent clip, fail with caption_no_speech and a next_action of use_overlay_captions. Neither is a policy rejection.

The probe call

The body is small: video_url (a media.sume.com artifact or asset in your workspace), frames false, and an Idempotency-Key. The default mode is sync, so the handler waits up to 30 seconds for a 200 and otherwise returns 202 with a job to poll. The result carries probe, and the docs name probe.has_audio as the field to read.

Inspect is billed by Modal compute, not at a fixed per-call price, and the docs give no number for a probe-only call. If you need a budget figure, read it from your usage after a trial call rather than assuming it is free.

curl -X POST https://api.sume.com/v1/video-inspect \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: probe-short-001" \
  -d '{"video_url": "https://media.sume.com/artifacts/artf_demo/short.mp4", "frames": false}'

Where the Shorts rules come in

The YouTube page makes a one-minute line important: a claimed Short over 60 seconds is blocked globally. Captions help with sound-off viewing but are unrelated to claims, so do not read the $0.20 job as a fix. The reason to probe is narrower: music-only AI clips, such as a generated video with no speech, are common, and they fail the speech path.

If has_audio is true but the audio is only music, the caption job can still end in caption_no_speech, so cues are the safer path for those clips. If you already hold the script, a script_text alignment keeps the speech timings and burns your text, which is a better fit for a voiced clip than cues are.

A short decision rule

Use three branches. If has_audio is false, skip speech-to-text and send cues with text, start and end. If has_audio is true and you know the clip is voiced, send the clip with a script_text if you have the script, or let speech-to-text run without one. If has_audio is true and the track is music or effects only, treat it as the silent case. In all three the price per accepted caption job is the same $0.20, so the probe is about avoiding a wasted attempt, not about changing the rate. Remember that cues and segments are the right tool for burned-in hooks and call-to-action lines too.