Can Any AI Video Model Pass This 7-Point Test?

0:00 Cling 4.0 just landed on Higsfield, so I gave it and eight other video models the exact same prompt. A brass robot barista pours a latte, draws a fox in the foam, slides the cup to the camera, and says

0:11 one line. Pick a number now. At the end, I'll tell you which model made it. First, what's actually new in Cling 4.0? The big one is length. One generation can now run 30 seconds in a single take.

0:23 Cling 3.0 stopped at 15. You can place up to 10 key frames across the shot. Cling 3.0 O only took a first and a last frame. It takes up to 15 references at once, 10 images, and five videos for

0:36 characters, products, or motion. It can also edit a clip you already have, and the sound is now stereo. That's Higsfield's spec sheet. My test asks something simpler. Does it follow the

0:46 prompt? To use it, open video on Higsfield, then create video, and pick Clling 4.0 in the model list. Set the length, the aspect ratio, and the quality. The cost shows on the generate

0:58 button before you click. Then there's this, the number of generations. I pushed it to four and Higsfield's own message said batch of four with no extra cost. The button still said 20 credits

1:09 and my credit history shows 20 credits for all four takes. So one click gave me four takes of the same prompt. Two of the four drew a clear fox in the foam. In the other two, I can't make one out.

1:20 Now listen to all four say the name. >> One flat white for Siobhan. One flat white for Siobhan. >> One flat white for Siobhan. One flat white for Siobhan. Four for four. It

1:32 said Shivon and Cling 4.0 got it every time. Keep that in mind because five of the other eight models got it wrong. On the seven checks, the four takes scored six, 6 and 1/2, 5 1/2, and 6 and 1/2.

1:46 Every take slid the cup and pushed in, but the steam is faint in all four. And watch take 2's arm at the end. It melts into smoke. Take four is the only one that cut to a close-up of the fox on its

1:57 own. But watch the cup across the cuts. The handle swaps sides and the fox turns back into a leaf. One flat white for Siobhan. Now the showdown. This is the exact prompt every

2:16 model got. Cinematic closeup in a cozy cafe at dawn. A small brass robot barista with round glass eyes pours steamed milk from a silver jug into a white cup, drawing a fox in the latte

2:26 foam. Steam rises through warm window light. The robot slides the cup across the wooden counter toward the camera and says in a calm, friendly voice, "One flat white for Siobhan, slow pushin,

2:38 shallow depth of field, realistic materials." I turned it into seven checks like unit tests. A model passes a check only if I can see it or hear it in the clip. For the name, I also ran every

2:48 clip's audio through a speech recognizer and offered it the right pronunciation and the wrong ones. The rules. Every model ran on Higsfield, which is sponsoring this video, with the same

2:58 prompt and the same settings, 10 seconds, 16x9, 1080p, sound on, first attempt only. Three things to say up front. Miniax H3 only renders in 2K. Four models were left on auto aspect

3:12 ratio, and all four came back 16x9 anyway. and Clling 3.0 had Higsfield's prompt enhancer switched on which I missed so its prompt may have been rewritten. That's the point of

3:23 Higsfield. One prompt box and every model is one click away in the list. The credits you'll see are what my plan was charged that day. Seedance 2.5 with its bit rate set to high. The priciest run

3:35 here. 160 credits. [music] [music] >> One flat white for Shibon. >> Five and a half. A clear fox, a real slide, and a steady push in. But the

3:52 name came out as Shibbon. The steam is faint. And watch the jug. It hangs in midair. Then it's suddenly standing on the counter. Seance 2.0. The version before it. Also on high bit rate 120

4:05 credits, [music] >> one flat [music] white for Shabban. >> Also five and a half. A big plume of steam and a fox. Though it looks like a

4:20 printed logo, but the slide is mostly the camera moving. It said Shabban. And at the end, its hands turn into flat discs. Miniax H3 in 2K 20 credits. >> [music]

4:38 >> One flat white for Shoen Four and a half. Its best moment is the slide. It really pulls the cup to the camera, but the foam is a plain tulip. There's no clear steam and the name came

4:50 out as Shoban Juan 3.0 35 credits. [music] One flat white for Siobhan. >> 6 and a half tied with cling 4.0's best takes. Clear steam, golden light, a real

5:10 slide, the name said right and a pushin once it gets going. The one miss the foam is a tulip with two ears so it gets half a point. Grock imagine 1.5 80 credits.

5:22 [music] One flat white for Siobhan. >> Five. Thick steam, a real slide, and the name said right. But the foam is a cat, not a fox, and the camera drifts

5:38 sideways instead of pushing in. Two glitches. The face is in the foam before the milk lands, and the cup grows a handle out of nowhere. Flux 3. 80 credits.

5:55 One flat white for Shouban. [music] >> Four. It drew one of the clearest foxes here, but there's no steam. The camera pushes in and then pulls back and the name came out as Shabban. Gemini Omni

6:07 Flash 1.1. 45 credits. One flat white for Seobin. Four and a half. A fox in profile. Real steam and golden light. But it read the name off the spelling. See Oon the camera pulls

6:28 back instead of pushing in. And most of the slide happens after its hands let go. And cling 3.0. 25 credits. One flat white for Siobhan. >> Five. It slid the cup, said the name

6:48 right, and pushed in. But it's gray daylight instead of a golden dawn. There's no fox. And a second jug fades in out of nowhere. And remember, its prompt enhancer was on. One version

6:59 apart, the two sea dance models tied at 5 and a half. 2.5 slid the cup better. 2.0 had more steam. Both got the name wrong. And 2.5 cost 40 credits more. Cling 3.0 scored five. Cling 4.0's 0's

7:13 four takes scored between 5 and 1/2 and 6 and 1/2. The difference, warm dawn light instead of gray daylight and a fox in two of the four takes where cling 3.0 drew none. Both said the name right.

7:26 Here's the whole test report. So the winners, no model passed all seven checks. WAN 3.0 and two of the cling 4.0 takes came closest at 6 and 1/2. For dialogue, cling 4.0 said the name right

7:39 in all four takes. And for value, its batch gave me four takes for 20 credits. So, which would I use? For a long detailed prompt, Juan 3.0 or Cling 4.0. For dialogue, cling 4.0. And when a

7:52 model offers a batch, use it. One of my four takes had a melting arm. Now, the grid from the start. Here's who made each number. Which one did you pick? And do your eyes agree with my checks? Tell

8:03 me in the comments. Every model in this video ran on Higsfield. The prompt is in the description. Paste it, switch models, and pick your own winner. The link is the first one there. This video

8:13 was sponsored by Higsfield.