Video API bake-off

Same prompt, three models, 16 September 2026. Paid from the $1 treg promo credit.

Short answer: subscribe to nothing

Every serious video model bills per second of output. Monthly plans are for people clicking in a web UI — they buy credits at a worse rate and expire them. For a pipeline, pay-as-you-go is cheaper and carries no commitment.

You already have the account that makes vendor signups optional: treg. One token reaches all of these, and the provider key is held server-side.

Winner on the test: Veo 3.1 Lite — $0.12 for 4 seconds, 1280×720, and the most photoreal of the three.

Seedance 1 Lite is the resolution bargain ($0.18 and it returned 1920×1088). Hailuo is the one that ships audio.

Dialogue — one pass, no separate voice track

Veo 3.1 Lite · a character speaking · sound on

$0.20

for 4s · 1280×720 · AAC 48 kHz stereo, mean −21.2 dB · the line, the voice, the lip movement and the room tone all came out of one generation — no TTS, no lip-sync pass

Download original

With audio

Veo 3.1 Lite · with sound · turn the volume up

$0.20

for 4s · 1280×720 · native synced audio · $0.05/s — same price per second as the 480p models, at 720p

Download original

The clips

Veo 3.1 Lite · Google

$0.12

for 4s · 1280×720 · silent (audio is $0.05/s vs $0.03/s) · ~90s to generate · cheapest per second of the three

Download original (2.8 MB)

Seedance 1 Lite · Replicate

$0.18

per 5s · asked for 720p, returned 1920×1088 · 5.04s · no audio · 40s to generate · $0.036/s

Download original (15 MB)

Hailuo H3-Max · MiniMax

$0.25

per 5s · 832×480 · 5.17s · has audio · ~60s to generate · $0.05/s (768P is $0.08/s = $0.40)

Download original (2.5 MB)

What a 5-second clip costs

ModelRate5s clipNote
Veo 3.1 Lite 720p silent$0.03/s$0.15durations 4, 6, 8 only
Seedance 1 Lite 720p$0.036/s$0.18480p = $0.09
Hailuo H3-Max 480P$0.05/s$0.25768P = $0.40
Wan 3.0 480p$0.05/s$0.25audio on by default
Veo 3.1 Lite 720p + audio$0.05/s$0.25cheapest good audio; $0.08/s is the 1080p row
Kling v3.0 Standard$0.084/s$0.42quality leader
Runway Gen-4.5$0.12/s$0.60
Seedance 2.5 720p$0.231/s$1.16reference-driven
Veo 3.1 full$0.2–0.6/s$1.00–3.004K, best realism
Sora 2 Pro$0.3–0.5/s$1.50–2.50

Quality ranking, September 2026

Artificial Analysis arena: Kling v3 is first at 1934, Alibaba Happy Horse 1.0 second at 1816, Seedance 2.0 Fast third at 1747. Veo 3.1 leads the with-audio table for realism and lip-sync. Hailuo is strong on physics. There is no single winner — the usual answer is to route two or three models by scene type, which one treg token makes cheap to do.

If volume becomes real

Open one account — Replicate, pay-as-you-go, cheapest per second and it carries Seedance, Kling, Veo, Wan and Hailuo under a single key. Register that key into treg and those calls stop touching the treg balance entirely, while the endpoint ids stay the same.