Kling 3.0 audio is on by default and adds 50% to the price

A video you did not need sound on can cost you half again as much. On Sume, kling-3 treats a missing generate_audio as true. The pricing code reads the field with a default of true and picks the audio-on rate, $0.168 a second at list, which Sume sells at $0.21. Pass generate_audio: false and the rate drops to $0.112 at list, or $0.14 at Sume. The difference is exactly 50%.

This matters in two places: batch jobs that never use the sound, and silent product loops that will be dubbed by a voiceover anyway.

The two requests

These are identical except for one field. The first costs $1.05 for 5 seconds, the second $0.70.

# audio on (the default when the field is omitted): 5 x $0.21 = $1.05
curl -s https://api.sume.com/v1/videos -H "Authorization: Bearer $SUME_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"model":"kling-3","prompt":"A paper boat on a puddle","duration":5,"resolution":"720p"}'

# audio off: 5 x $0.14 = $0.70
curl -s https://api.sume.com/v1/videos -H "Authorization: Bearer $SUME_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"model":"kling-3","prompt":"A paper boat on a puddle","duration":5,"resolution":"720p","generate_audio":false}'

What it adds up to

For a catalog of 200 five-second product loops, leaving audio on is $210 and switching it off is $140, a $70 difference for sound you would then mute. At 15 seconds the gap per clip is $1.05.

Reading the bill

You can confirm which rate applied without doing any arithmetic. The completed poll response carries usage.cost as the Sume billable amount, so a 5-second 720p Kling job that shows 1.05 was billed with audio on and one that shows 0.7 was billed with audio off. If you find 1.05 where you expected 0.70, look at your request body first: a client library that drops a false-valued field, or a template that builds the JSON without the key, will quietly fall back to the default.

A good habit is to send the field explicitly in every request, true or false, so the choice is visible in your logs and in code review, and to assert on usage.cost in a smoke test.

Not every model works this way

The audio control differs by model, so do not copy the field across ids blindly. Seedance, Wan and Kling take generate_audio as an option. Gemini Omni Flash 1.1, MiniMax H3 and H3 Max always render native audio and the field should be omitted. On sume/auto, which resolves to Omni, sending generate_audio: false fails with 400 unsupported_capability.

If you want a silent clip from Omni or H3, drop the audio track in an edit step after the fact. If you want it from Kling, the flag is enough, and the saving shows up as the lower usage.cost on the completed job.