AI video with native audio: which Sume models let you turn it off

Among Sume video models, Kling Video v3 Pro is the one with an audio switch that changes the price: $0.14 per second with audio off and $0.21 with audio on. Gemini Omni Flash 1.1, MiniMax H3 and MiniMax H3 Max always produce native audio and reject a request to turn it off. Grok Imagine Video 1.5 and Genjutsu produce none. Read from the Sume catalog, as of 2026-10-08.

Google's Gemini API video page presents Veo 3.1 as a native-audio video model (Google page, read 2026-10-08); Sume does not list Veo.

Why the toggle matters for Kling

A 10-second Kling clip is $1.40 with audio off and $2.10 with audio on, a difference of $0.70. If you will lay your own music or a voice-over track on top, pay the lower rate. The fal page for Kling v3 Pro states the same two per-second rates, $0.112 and $0.168 (fal page, read 2026-10-08).

Because Sume's default estimate for kling-3 assumes audio on ($1.05 for a 5-second clip), send generate_audio: false explicitly when you want the cheaper rate.

When audio is not optional

For Gemini Omni Flash 1.1 and MiniMax you pay for audio inside the per-second rate, so there is nothing to save by trying to disable it. If you need a silent clip, remove the audio track in your editor or with the audio-detach workflow, and do not expect the model to skip it.

A 5-second Gemini Omni clip at 720p is $0.625, and a 5-second H3 clip at 768p is $0.375; both include the audio.

Cost of leaving audio on by default

Over a month of 1,000 Kling clips at 10 seconds each, leaving audio on costs 1,000 x $2.10 = $2,100, while turning it off costs 1,000 x $1.40 = $1,400, a saving of $700 when you do not use the generated audio. That is the only video model in the catalog where the toggle moves the bill.

For the other models, price is a function of resolution and length only, so the lever is the resolution tier.

Practical checklist

• Decide whether the audio will be used before you pick Kling.
• Check generate_audio in GET /v1/videos/models for each model; it shows the audio capability.
• Do not send generate_audio on the Video 1.0 route: the Auto family always generates audio.
• For timing-critical voice, generate the video silent or with reference audio and mix the voice separately.