Does audio off make AI video cheaper? Kling 3 yes, Wan 3.0 no

Setting generate_audio to false makes a clip cheaper on exactly one of Sume's main video rows: Kling 3, where a silent second costs $0.14 and a second with sound costs $0.21. On Wan 3.0 the flag does not move the price at all, and on Seedance 2.5, MiniMax H3 and Gemini Omni Flash 1.1 sound is either not priced separately or cannot be switched off.

That matters when you budget a batch, because the cheap-looking habit of sending generate_audio: false everywhere saves money in one place and only removes sound everywhere else. The figures below come from Sume's price tables in the repository (provider list times the 1.25 house margin, rounded up to the cent per job) and the Video Generation page, read 2026-10-10.

Which rows put a price on sound

Kling 3 has two per-second rates, and the audio-on rate is the default when you omit the field. Wan 3.0 reads the flag but prices by resolution only. Gemini Omni Flash 1.1 always produces synced audio and has no audio rate. MiniMax H3 and H3 Max always produce native stereo audio, and the Video Router schema refuses generate_audio: false for them with a message to omit the field. Seedance 2.5 bills video tokens, so there is no sound line to switch off.

What it does to a real clip

Take a 5 second Kling 3 clip. With sound it reserves and bills $1.05; silent it is $0.70. At 10 seconds the pair is $2.10 and $1.40. A hundred 10 second clips therefore differ by $70.00 between the two settings, which is the only place in this comparison where the flag is a budget lever.

For Wan 3.0 the same 5 second clip is $0.32 at 480p, $0.63 at 720p and $1.25 at 1080p whether or not you ask for audio. If you do not need sound on Wan, turning it off saves nothing, so decide on creative grounds. Because the estimator rounds up per job, those cents come from $0.3125, $0.625 and $1.25 before rounding.

Why the default can surprise a budget

Kling's audio default is on. A request that omits generate_audio is priced at $0.21 per second, so a spreadsheet built from the silent rate underestimates every omitted-field call by a third. When you build a cost preview, read the field you will actually send rather than assuming silence, and write the value explicitly in the request so the number you computed matches the number that is reserved.

Reservation follows the same math. At submit Sume holds the estimated amount, and a balance that is short of that amount returns 402 insufficient_credits before any provider work starts, as described in Generation admission. A 15 second Kling clip with sound needs $3.15 free in the wallet; silent it needs $2.10.

A short rule for choosing

Use the silent Kling rate when you will lay a music or voice track over the clip afterwards and do not need the model's own sound. Keep audio on when ambient sound or speech is part of what you are buying. On every other row, ignore the flag when you estimate cost.

Two cautions apply. First, Veo 3.1 prices audio on and off separately in Sume's tables, but the Veo rows only appear where the Google Cloud route is configured, so check GET /v1/videos/models before you plan around them. Second, rate cards can change; the catalog and the Errors and rate limits page describe what happens when a balance is short, but the per-second numbers belong to the live catalog, so re-read it before a large batch.

• Kling 3: audio off saves $0.07 per second.
• Wan 3.0, Omni, H3, H3 Max, Seedance 2.5: no saving from the flag.
• Always send generate_audio explicitly when your estimate depends on it.