How to pick an AI video model: run a bake-off on your own shots

To pick an AI video model, run a bake-off. Choose the two or three shots you post most, make one still per shot, and run the same prompt at the same length on every model you are thinking about. Score each take on one sheet: did it pass the realism check, did it do what the prompt says in the right order, does the mouth match the voice, what did it invent, and what did it cost. The model that wins your shot is your model for that shot. A leaderboard or a demo reel cannot tell you that, because neither one ran your shot.

We ran one on November 14, 2026, in Hook Studio, an AI reel maker for TikTok and Instagram: a talking shot, a product shot, and a shot with three actions in a set order, on Kling 3, Kling 4, Seedance 2, and Seedance 2.5. Fourteen clips, one take each, 2,126 credits with the stills. No model passed every shot. Every take is on this page, failures included. The people are AI people from Hook Studio's premade cast. For what each model is good at, see which AI video model to use for each job.

Every model is good at some shots and bad at others, and the only way to find out which is which for your shots is to run them. Public leaderboards help you pick candidates, not winners. The Artificial Analysis video arena ranks models by blind votes on short clips. When we read its image-to-video board on November 14, 2026, Seedance 2.5 was first and Kling 3.0 was 23rd. On our talking shot the order flipped: Kling 3's takes passed in English and Spanish, and both Seedance 2.5 takes failed. A vote on someone else's clip does not score your prompt, your person, or your price.

The lineup also changes under you. Google shut down its Veo 3.1 models in the Gemini API on October 22, 2026, and points developers to Gemini Omni 1.2 Flash (Google's deprecation list). OpenAI closed the Sora app in April 2026 and its API in September (OpenAI). A bake-off is not a one-time verdict. It is a test you run again when a model you use goes away or a new one comes out.

1. 1Pick the two or three shots you post most. A talking hook, a product in someone's hands, a shot where a person does a few things in order. Test those, not a scene you will never post.
2. 2Make one still per shot and approve it before any clip. Every model starts from the same first frame, so the still is your control. A still costs 11 credits or less in Hook Studio; one of ours took four tries.
3. 3Write one clip prompt per shot and paste it unchanged into every model. Name the action, the order, the camera, and any spoken words in quotes. If you tweak the prompt for one model, you are testing the prompt, not the model.
4. 4Run the same length on every model. We used 5 seconds. Write down the resolution too: Kling 3 renders 1080p only, so it ran at 1080p and the others at 720p.
5. 5Score every take on the same sheet. Watch it muted, then with sound. Use the five questions below.
6. 6Treat one take as a sample. Pick a finalist or two per shot and run two or three more takes of each before you commit. A model can pass once and fail the next take.
7. 7Keep a log of which model wins which shot. Most of your budget goes to remaking your winners on the model that won. A smaller share tests new models and formats, and that is where the next bake-off comes from. That split is the 70/30 rule.

Shot: ________  Model: ________  Resolution: ____  Length: __ s  Take: __
Price: ____ credits

1. Realism check: passed / failed. What it flagged: ________
2. Did it do what the prompt says, in order? yes / partly / no. What it skipped: ________
3. Lip sync (talking shots): matched / out of step from __ s to __ s
4. What it invented or changed (people, props, text, camera): ________
5. Watched muted, then with sound. Would I post it as is? yes / no

Verdict: finalist / out

The realism check answers question 1 in Hook Studio: it flags what looks fake in each still and clip. It does not answer question 2. Whether the clip did what you asked, in the order you asked, you judge from the frames yourself. In our test, the check passed a Kling 4 take in which the painting on the wall changed. Once a model wins, the winning format decides what you remake with it.

We ran it in Hook Studio on November 14, 2026, in the organic style, which looks shot on a phone by someone who was there. Every clip is 5 seconds, starts from the same approved still for its shot, and uses the same prompt for its shot, word for word. Kling 3 renders 1080p only; the other three models ran at 720p. The talking shot ran on three models: Kling 3, Kling 4, and Seedance 2.5. The product and direction shots ran on all four. One take per model per shot, so read every verdict below as a sample, not a final ranking. Hook Studio is not affiliated with Kling (Kuaishou) or Seedance (ByteDance).

Camila sits at a small table by the window of a Mexico City café with a cup of coffee and talks to her friend across the table.

Here's the still. The realism check passed it, with one note: it's framed to the waist, not the chest.

Still · 11 credits

Make her say: "Same photo, same line, different AI model. Which one looks real?" Run it on Kling 3, Kling 4, and Seedance 2.5.

Three clips, each waiting for your approval. The first:

5 s clip with speech · Kling 3 · 1080p · 143 creditsApprove

The Kling 3 take passed the realism check, and her mouth matches the voice in both sentences. The Kling 4 and Seedance 2.5 takes failed: both over-act the line.

Step one: the stills, and the three that failed first

The talking and product stills passed on the first try. The direction still took four. We first put Nia outside a bakery, and the realism check failed it for readable made-up words on the awning, the door, a chalkboard, and the price tags. Two more tries outside failed because the storefront and then a brick wall ran off at an angle behind her, and we had asked for a straight-on view. Moving her indoors, in front of a plain wall with no signs, fixed it. Six stills, 64 credits, before a single clip ran.

Shot 1: talking to camera, in English and Spanish

Camila says one line to a friend across the table, in the clip's own voice. The Spanish takes use the same prompt with "in Mexican Spanish" added and the line "Misma foto, misma frase, otro modelo de IA. ¿Cuál se ve real?" Turn the sound on to hear each one.

Seated at the café table, she is already talking to her friend across the table, calm and friendly, with a small natural nod; her hand stays near the cup. She says: "Same photo, same line, different AI model. Which one looks real?"

Camera: her friend, sitting across the table, eye level, handheld phone, slight natural hand sway. 5 seconds, with speech.

The same shot in Spanish:

Kling 3 was the only model whose talking takes passed, in both languages. Kling 4 and Seedance 2.5 over-acted the line. Kling 4 kept the mouth in time with the voice in both languages; Seedance 2.5 drifted in both. The Spanish Kling 3 take passed the realism check and still ran its voice half a second early at the end of the line. Hook Studio offered a lip-sync repair for 71 credits. Six talking takes cost 948 credits.

Shot 2: a product in someone's hands

Marco pours coffee from a glass carafe into a mug, then sets the carafe down. Liquid, glass, and two hands are hard for any video model, which is why it is in the test.

He is already pouring: the coffee streams from the carafe into the green mug and the mug fills about halfway; then he sets the carafe down on the counter beside the mug. Calm, unhurried, no talking.

Camera: his roommate, standing across the counter, chest height, handheld phone, slight natural hand sway. 5 seconds, no sound.

Kling 4 and Seedance 2.5 passed. Kling 3 did the pour and the set-down, and invented a second man who looks in from the right edge, then leaves an arm behind. Seedance 2 made the most believable coffee of the four, and failed because the camera pushed in until his face was cut off. Four takes, 557 credits.

Shot 3: three actions, in order

This is the direction test. Nia does three things in a set order in 5 seconds: sip, set the mug on the shelf, wave with both hands. A model that follows direction does all three, in that order, without breaking the room.

Three actions, in this order, starting at once: first she takes one sip from the white mug; then she steps toward the narrow wooden wall shelf at waist height and sets the mug down on it; then she turns back to the camera and waves with both hands, smiling. No talking.

Camera: her friend, standing about two meters in front of her in the living room, eye level, handheld phone. 5 seconds, no sound.

All four models did the three actions in the right order. What separated them was the room. Kling 3 failed: the narrow shelf grew into a long one, and the mug she put on it was gone by the end. The other three passed. Seedance 2.5 had the clearest wave. Seedance 2 did everything, with a small wave and her hands low. Kling 4 changed the painting on the wall, which the realism check did not flag. Four takes, 557 credits.

No model passed every shot. The model that won the talking shot failed both of the others. That is the point of running your own bake-off: a winner is a winner for one kind of shot.

What it cost

Credits are Hook Studio's prices on November 14, 2026, for 5 second clips, and the failed takes were charged like the others. Compare the cheapest take that passed, not the cheapest clip: 143 credits for talking (Kling 3), 134 for the product shot (Kling 4), and 130 for the direction shot (Seedance 2).

What goes to round two

One take per model gives you leads, not verdicts. From this round, the finalists are Kling 3 for talking, Kling 4 and Seedance 2.5 for the product shot, and Seedance 2.5 and Seedance 2 for the direction shot. Kling 4 passed the direction shot too, but it changed the painting, so it sits out. Each finalist gets two or three more takes of the same shot, on the same still and prompt, before it goes into the log as the winner.

• Winners are per shot. Kling 3 won the talking shot in both languages and failed the product and direction shots. Log the winner next to the shot, not next to your account.
• The realism check flags what looks fake, not what is wrong. It passed a Kling 4 take in which the painting changed. Check every take against the prompt yourself, frame by frame.
• When every model changes the same thing, look at the still. Our still showed the shelf with one strap. All four models rebuilt it as a longer shelf with two straps; the check failed only Kling 3 for it. Show a prop whole in the still, or leave it out.
• Keep words out of the still. Our first direction still failed for made-up text on a bakery's awning, door, chalkboard, and price tags. A plain wall fixed it. Put text on screen as an overlay instead.
• Test each language you post in. Kling 3's English take matched the voice; its Spanish take ran the voice 0.5 s early at the end. Kling 4 failed for different reasons in each language.
• Price the pass, not the clip. The cheapest clip in the test, Kling 3 at 96 credits, failed both silent shots. The cheapest take that passed the product shot cost 134.

• It does not score direction. The realism check judges what looks real. Whether the clip did what the prompt asks, in order, is your call.
• There is no bake-off button. Each clip is its own job, with its own model and its own price to approve.
• Failed takes cost the same as passed ones. Our three failed stills and seven failed clips were all charged.

For what each model is known for, beyond one bake-off, see which AI video model to use for each job, realism vs control, and Kling vs Seedance. For how to make the still and the clip in the first place, see how to make AI videos.