Native audio or separate voiceover?

The premium, quantified

Native-audio models (Veo 3.1, Sora 2) price sound into every second — including the 40–60% of seconds you throw away. A silent model plus separate voiceover bills audio exactly once, on the final cut. At a ×2 retake ratio, native audio's effective sound premium doubles.

When native audio wins anyway

• Lip-synced dialogue — the case native audio exists for; no post workflow matches Veo's sync.
• Timing-critical explainers — generating word and image together beats aligning them later.
• Diegetic sound — footsteps, impacts, ambience that must match on-screen action.

When separate audio wins

• Voiceover-driven content (most of YouTube): narration recorded once over silent footage — Seedance + a TTS/voice subscription beats any native-audio stack on cost.
• Music-bed content: you'll mute generated audio anyway — paying for it is pure waste.
• Multi-language: one silent master, many voice tracks.

Decision rule: does the mouth move on camera? Yes → pay the native premium on those shots only. No → silent tiers plus one audio pass. The mixed routing is standard practice, not a hack.