Research

AI Video Model Benchmark 2026: 19 models, one prompt, measured outputs

On 2026-07-13 we ran 23 real generations across 19 AI video models (VEO 3.1, Seedance 2.0, the Kling family, Hailuo 2.3, PixVerse V6 and more) with one identical prompt, then probed every delivered file with ffmpeg. The table below reports what the models actually returned: measured resolution, measured duration, whether an audio track exists, and the historical cost at the test date. Republish the table with attribution to this page.

real generations: 23
AI video models: 19
identical prompt: 1
files with an audio track: 16 of 23

Historical delivery test, not a quality leaderboard. Generations: 2026-07-13. Methodology and wording reviewed: 2026-09-18. The July table remains historical. A separate six-attempt September delivery pilot is included below with its own requests and files.

These 23 outputs use different configurations and one scene. They do not measure repeatability, visual quality, failure rate, generation speed or customer sales. An audio track does not establish intelligible speech. Source media and provider request logs are not included in this public table, so the full experiment cannot be independently reproduced from these rows alone.

Download the 23 recorded results (CSV)Download the next-test protocol and worksheet Success rates and wait times from real use Check current credit packs

New measured examples · 18 September 2026

Three scenes. Two attempts each. Every result.

We ran six new text-to-video jobs with VEO 3.1 Lite: 4 seconds, 720p, vertical 9:16, audio off. These are original prompts, with no uploaded photograph. The fictional bottle is an imagined product; this test does not establish that a real product label will be preserved.

6 of 6 attempts met the file checks. We checked downloaded dimensions, duration and audio streams, and recorded every attempt below. Total net debit: 18 credits from a marked test account. This is a small delivery pilot, not a model ranking, reliability guarantee or sales experiment.

Requests & results (JSON)Measured results (CSV)

A fictional product label

A single amber glass bottle with a simple white label reading ALTO stands on a warm white studio pedestal. Slow camera push toward the bottle, gentle light moving across the label. Preserve its rigid shape and exact lettering. Minimal product commercial, no people, no additional text.

Attempt 1 · 3 credits

720×1280 · 4s · No audio track

ALTO is readable in sampled frames. This attempt produced an uncapped bottle; the second has a cap.

Download attempt 1

Attempt 2 · 3 credits

720×1280 · 4s · No audio track

ALTO is readable in sampled frames. Bottle, cap and label proportions differ from the first attempt.

Download attempt 2

Try this promptOpens a draft. Review settings and cost before generating.

A rigid object in motion

A single rigid red cube rests on a matte blue studio platform with a matching seamless background. The camera makes a slow precise quarter orbit. Keep all six faces straight and the cube shape unchanged. Crisp editorial lighting, no people, no text.

Attempt 1 · 3 credits

720×1280 · 4s · No audio track

Straight cube edges remain visible in sampled frames. A circular display base was generated.

Download attempt 1

Attempt 2 · 3 credits

720×1280 · 4s · No audio track

Straight cube edges remain visible in sampled frames. The display base is square rather than circular.

Download attempt 2

Try this promptOpens a draft. Review settings and cost before generating.

A coastal scene

A quiet rocky coastline at sunrise, gentle waves washing around one foreground rock. Slow steady camera push toward the ocean, natural golden light, clear horizon, no people, no text.

Attempt 1 · 3 credits

720×1280 · 4s · No audio track

The sampled frames retain the foreground rock and horizon while the water changes.

Download attempt 1

Attempt 2 · 3 credits

720×1280 · 4s · No audio track

The sampled frames retain the rock and horizon; composition and wave patterns differ from the first attempt.

Download attempt 2

Try this promptOpens a draft. Review settings and cost before generating.

Method and limits

Protocol written before submission: three original prompts, two attempts per prompt, fixed interleaved order, maximum six submissions and 18 credits. Same model and settings throughout; no seed fixed. All attempts are included. No private customer media was used.

File pass means a downloadable MP4 with a 720×1280 video stream, duration within 0.2 seconds of 4 seconds and no audio stream. Files were measured with ffprobe; SHA-256 hashes, exact requests and timestamps are in the JSON. Completion was observed by polling, so elapsed time includes network and polling delay.

Visual notes are one AI-assisted review of sampled frames, not an independent blind panel or a full temporal-quality score. No inference about ad performance, purchased-product accuracy, other settings or other models is supported. The model can return different results on another attempt.

Credit debits and refunds were checked on the isolated test account. No pack was purchased for this pilot, so no actual paid USD-per-credit or provider invoice cost is claimed. Current prices may change; check the editor before submitting.

July findings

Cheapest real clip: Hailuo 2.3 at 3 credits (~$0.54) — but its standard tier outputs landscape 768p regardless of the aspect request.
Lowest-credit vertical clip with an audio track in these rows: PixVerse V6 — 720×1280 with an audio track at 4 credits (~$0.72).
Cheapest native-audio 1080p: VEO 3.1 Lite at 6 credits for 8 seconds (1920×1080, Google-generated audio).
Kling 3.0 Pro over-delivered: a 720p request returned true 1080×1920.
Kling 2.5 Turbo Pro and LTX-2 19B returned landscape/4:3 files for the recorded vertical requests. This test did not isolate provider behavior from request mapping.
Hailuo 2.3 Pro (1080p) is fixed at ~6 seconds — a 10-second request cannot be fulfilled, so Viralance caps the price to the 6s rate automatically.
Silent toggles are real: Kling 3.0 Standard and Kling O3 Pro files contained no audio track when silent was selected (O3 Pro is also discounted from 12 to 10 credits when silent).
Every delivered duration was within 0.1s of the request except the fixed-length cases noted above.

July results

Prompt used for every run: “A golden retriever puppy running through a sunny meadow, cinematic natural light”. USD column uses the historical largest-pack rate from this test ($0.181/credit); smaller packs cost more per credit.

Historical July 13, 2026 generation outputs and recorded costs
Model	Requested	Credits	~USD	Measured output	Measured length	Note
VEO 3.1 FastGoogle DeepMind	T2V · 6s · 720p · 9:16	8	$1.45	720×1280	6.0s	—
VEO 3.1 FastGoogle DeepMind	I2V · 8s · 1080p · 9:16	16	$2.90	1080×1920	8.0s	VEO image-to-video is fixed at 8s
VEO 3 FastGoogle DeepMind	T2V · 4s · 720p · 9:16	6	$1.09	720×1280	4.0s	—
VEO 3.1 LiteGoogle DeepMind	T2V · 8s · 1080p · 16:9	6	$1.09	1920×1080	8.0s	Cheapest native-audio 1080p in the test
Seedance 2.0ByteDance	T2V · 5s · 720p · 9:16	12	$2.17	720×1280	5.1s	—
Seedance 2.0ByteDance	T2V · 10s · 1080p · 16:9	55	$9.97	1920×1080	10.1s	Most expensive configuration tested
Kling 3.0 ProKuaishou	T2V · 10s · 9:16	18	$3.26	1080×1920	10.0s	Delivered 1080p on a 720p request — free upgrade
Kling 3.0 StandardKuaishou	T2V · 5s · 1:1 · silent	8	$1.45	960×960	5.0s	Silent toggle honored — no audio track in file
Kling O3 ProKuaishou	T2V · 5s · 1080p · silent	10	$1.81	1080×1920	5.0s	Silent discount applied (10cr instead of 12)
Kling 2.5 Turbo ProKuaishou	T2V · 5s · 9:16	4	$0.72	1920×1080	5.0s	Recorded 9:16 request; delivered landscape (request mapping not isolated)
Grok ImaginexAI	T2V · 6s · 720p · 9:16	4	$0.72	720×1280	6.0s	—
Vidu Q3 TurboVidu	T2V · 8s · 1080p · 16:9	6	$1.09	1920×1080	8.0s	—
LTX-2 19BLightricks	T2V · 5s · 9:16	4	$0.72	1024×768	4.8s	This run delivered 1024×768 for the recorded 9:16 request
WAN 2.7Alibaba	T2V · 5s · 9:16	5	$0.91	720×1280	5.0s	—
Hailuo 2.3MiniMax	T2V · 6s · 720p · 9:16	3	$0.54	1366×768	5.9s	Cheapest clip in the test; standard tier outputs landscape 768p (no aspect parameter)
Hailuo 2.3 (Pro)MiniMax	T2V · 1080p	4	$0.72	1920×1080	5.9s	Pro (1080p) tier is fixed-length ~6s — Viralance caps it to 6s pricing automatically
Happy Horse 1.1Alibaba	T2V · 10s · 1080p · 9:16	19	$3.44	1080×1920	10.1s	Audio track present; lip-sync quality was not assessed
PixVerse V6PixVerse	T2V · 5s · 720p · 9:16	4	$0.72	720×1280	5.0s	Vertical output with an audio track at 4 credits in this run
PixVerse V6PixVerse	I2V · 8s · 1080p · 9:16	10	$1.81	1072×1920	8.0s	—
PixVerse V5PixVerse	I2V · 5s · 720p · 9:16	6	$1.09	720×1280	5.0s	Image-to-video only
Kling 2.6 ProKuaishou	I2V · 5s · 9:16	6	$1.09	1088×1904	5.0s	Image-to-video only; audio generated
Hunyuan 1.5Tencent	I2V · 5s · 9:16	4	$0.72	480×848	5.0s	Outputs 480p only — labeled honestly on Viralance
Grok Imagine 1.5xAI	I2V · 5s · 720p · 9:16	5	$0.91	720×1264	5.0s	Image-to-video only