Benchmark · v2026.10

AI video model benchmark: one photo, seven models.

The same still and the same prompt, sent to Veo 3.1, Kling V3 Pro, Kling O3 Pro, Seedance 2.0, WAN 3.0, MiniMax H3 and LTX 2.5 Pro as image to video. Each model's real clip, what that clip cost, and how long the model takes on real jobs.

Version 2026.10 · published 7 October 2026 · next update November 2026

Download the CSV Embed the table

The test

One still, one prompt.

The shared starting still: a paper boat in a flooded cathedral at dawn

Prompt “A paper boat the size of a ship sails through a flooded cathedral at dawn, candle flames guttering as it passes, water lapping the stone, slow lateral dolly, 35mm, natural light, the sound of water and wind, continuous motion, no cuts, no new objects”

Results

Seven models, side by side.

Sort by any column. Dashes mean the figure could not be measured honestly; see the method.


Veo 3.1	Google DeepMind	1080p	8 s	Yes	4	32	$12.79	$8.90	112 s (108–124, 12 jobs)8 s clips: 134 s (3 jobs)
Kling O3 Pro	Kuaishou (Kling)	1080p	8 s	Yes	1.4	11.2	$4.48	$3.12	122 s (106–174, 163 jobs)
Kling V3 Pro	Kuaishou (Kling)	1080p	8 s	Yes	1.68	13.44	$5.37	$3.74	128 s (113–158, 92 jobs)
Seedance 2.0	ByteDance	1080p	8 s	Yes	6.804	54.43	$21.75	$15.14	287 s (265–312, 20 jobs)
WAN 3.0 1080p	Alibaba	1080p	8 s	Yes	2	16	$6.39	$4.45	—
MiniMax H3 2K	MiniMax	2K	10 s	Yes	1.3	13	$5.19	$3.62	185 s (132–264, 305 jobs)10 s clips: 201 s (66 jobs)
LTX 2.5 Pro	Lightricks	1080p	8 s	Yes	1.7	13.6	$5.43	$3.78	—

Method

How the numbers were made.

Every model received the same still (made once on Nano Banana Pro, 16:9) and the same prompt, as image to video, at the clip length shown. The clips were rendered on 3 September 2026 and are shown as each model delivered them, re-encoded for the web. Click a frame to play its clip.

Prices are the credits King AI's catalog bills for that exact clip (catalog v119), and the dollar figures convert them at the $9.99 pack's rate ($0.3996 a credit, the most a credit costs) and the Pro plan's ($0.2782, the least).

Generation time is not timed on these clips; it is the median of real production jobs for the same model, measured 7 October 2026. From the moment King AI's server submits the job to fal until fal has the finished video. It includes fal's queue (median 0.1 seconds) and excludes King AI's own step of storing the video and making its thumbnail, which adds a median of 3 to 6 seconds. Most of those jobs were shorter clips than the ones shown here, so the per-length figure is given where there were enough jobs at the take's length.

WAN 3.0 1080p: 13 jobs spread from 5 to 30 seconds long, so one median would mislead; clips of 5 seconds took a median of 145 seconds (5 jobs).
LTX 2.5 Pro: Only 2 production jobs in 90 days (41 and 73 seconds); too few for a median.

Reuse

Embed or republish the table.

Free to republish or embed, with a visible link to kingaivideo.app/benchmarks. Paste the HTML below into your page; the caption carries the link. Or download the CSV.

<div style="overflow-x:auto"><table><caption>AI video model benchmark v2026.10 &middot; source: <a href="https://kingaivideo.app/benchmarks">King AI Video benchmark</a></caption><thead><tr><th>Model</th><th>Maker</th><th>Resolution</th><th>Clip</th><th>Audio</th><th>Credits / s</th><th>Credits, this clip</th><th>Cost, $9.99 pack</th><th>Cost, Pro plan</th><th>Median generation time</th></tr></thead><tbody><tr><td><a href="https://kingaivideo.app/models/veo-3">Veo 3.1</a></td><td>Google DeepMind</td><td>1080p</td><td>8 s</td><td>Yes</td><td>4</td><td>32</td><td>$12.79</td><td>$8.90</td><td>112 s (12 jobs)</td></tr><tr><td><a href="https://kingaivideo.app/models/kling">Kling O3 Pro</a></td><td>Kuaishou (Kling)</td><td>1080p</td><td>8 s</td><td>Yes</td><td>1.4</td><td>11.2</td><td>$4.48</td><td>$3.12</td><td>122 s (163 jobs)</td></tr><tr><td><a href="https://kingaivideo.app/models/kling">Kling V3 Pro</a></td><td>Kuaishou (Kling)</td><td>1080p</td><td>8 s</td><td>Yes</td><td>1.68</td><td>13.44</td><td>$5.37</td><td>$3.74</td><td>128 s (92 jobs)</td></tr><tr><td><a href="https://kingaivideo.app/models/seedance">Seedance 2.0</a></td><td>ByteDance</td><td>1080p</td><td>8 s</td><td>Yes</td><td>6.804</td><td>54.43</td><td>$21.75</td><td>$15.14</td><td>287 s (20 jobs)</td></tr><tr><td><a href="https://kingaivideo.app/models/wan-3">WAN 3.0 1080p</a></td><td>Alibaba</td><td>1080p</td><td>8 s</td><td>Yes</td><td>2</td><td>16</td><td>$6.39</td><td>$4.45</td><td>—</td></tr><tr><td><a href="https://kingaivideo.app/models/hailuo">MiniMax H3 2K</a></td><td>MiniMax</td><td>2K</td><td>10 s</td><td>Yes</td><td>1.3</td><td>13</td><td>$5.19</td><td>$3.62</td><td>185 s (305 jobs)</td></tr><tr><td><a href="https://kingaivideo.app/models/ltx-2-5">LTX 2.5 Pro</a></td><td>Lightricks</td><td>1080p</td><td>8 s</td><td>Yes</td><td>1.7</td><td>13.6</td><td>$5.43</td><td>$3.78</td><td>—</td></tr></tbody></table></div>

FAQ

Benchmark questions.

Which AI video model is fastest?

Of the models with enough production jobs to measure, Veo 3.1 had the shortest median generation time on real jobs (112 seconds) and Seedance 2.0 the longest (287 seconds). Those jobs were mostly clips of 4 to 5 seconds, and longer clips take longer. Times include the model provider's queue, which was well under a second.

Which model cost the least for this clip?

Kling O3 Pro, at 11.2 credits for its 8-second take ($4.48 at the $9.99 pack's rate). Seedance 2.0 cost the most, at 54.43 credits. Prices are per second, so clip length matters as much as the model.

Can I use this data?

Yes. You may republish or embed the table and the CSV, free, with a visible link to kingaivideo.app/benchmarks.

How often is the benchmark updated?

Monthly. This is version 2026.10, published 7 October 2026; the next update is due in November 2026, with prices re-read from the catalog and generation times re-measured.

Keep exploring