We paid $2.47 to test six talking-video models. Here is the table.
Six image-to-video models with native speech, tested on the same two start frames: price per 5 s clip, generation time, lip sync, what broke, and our pick.
The usual way to pick an AI video model is to watch each vendor's showreel and choose the prettiest. The showreels are not the same prompt, not the same start frame and not the same budget, so they tell you nothing about the choice you actually face: a person in a still image has to say a specific sentence to camera, with their own voice, and the clip has to cost less than a coffee.
So we ran that exact test. Two start frames, six models, the same scripted line per frame, native audio on, through one API. Twelve clips, twelve delivered, no failures, no retries, and a total bill of 493 credits, which is $2.47. Which model said the words, which one kept the person's face, which one melted the prop, and which two would we actually use?
The setup
Both start frames were stills from our UGC pipeline, generated to look like a phone's front camera with the phone outside the frame. Set A: a founder at a desk holding a cold brew. Set B: a welder on a shop floor holding a machined steel block. Each model got the same prompt per set: the spoken line in quotes, one mid-action (a sip, lifting the block once), a background sound, and a fixed suffix asking for eye contact, handheld shake, no captions, no text, no camera UI, no logos.
All calls went through KIE's API, where 1 credit is $0.005, at 720p with audio generation on. Five models were asked for 5 seconds; Veo 3.1 Fast produces 8 second clips. The test was run on 13 October 2026.
The price table
| Model | Per clip | Clip length | Generation time | Per 20 s reel |
|---|---|---|---|---|
| Grok Imagine 1.5 | 22.5 credits, $0.11 | 5 s | 31 to 69 s | $0.45 |
| Seedance 1.5 Pro | 35 credits, $0.175 | 5 s | 44 to 50 s | $0.70 |
| MiniMax H3 | 40 credits, $0.20 | 5.2 s | 100 to 118 s | $0.80 |
| Seedance 2.0 Mini | 41 credits, $0.205 | 5.1 s | 87 to 112 s | $0.82 |
| Pixverse V6 | 48 credits, $0.24 | 5 s | 45 s | $0.96 |
| Veo 3.1 Fast | 60 credits, $0.30 | 8 s | 81 to 100 s | $0.90 (24 s) |
The "per reel" column assumes four 5 second clips, or three 8 second clips for Veo. Every model lands under $1 of clips per reel. Price alone does not separate them much; the results do.
What each model did with the founder
All six spoke the full line and moved the lips on the words. The differences were in the acting, the face and the prop.
Grok 1.5 sipped through the straw and put the glass back; loud, clear audio at -12.1 LUFS; the brand name came out as "ultimate". Seedance 1.5 Pro got the brand name right, sipped plausibly, but had half-closed eyes in the first second. Seedance 2.0 Mini drank from the rim, and the window view brightened while the laptop shifted: scene drift. MiniMax H3 was fine but soft (a 0.8 MB file at low bitrate) with audio peaking at 0 dBFS. Pixverse V6 smoothed the face into slightly plastic skin. Veo 3.1 Fast gave the most natural performance of the twelve: a sip, the glass set down, then a hand gesture to camera.
What each model did with the welder
The welder clip is the prop test, because a block of steel has a shape that should not change. Grok 1.5 lowered the block instead of lifting it, and by 3 to 5 seconds it had shrunk into a small cylinder. Seedance 1.5 Pro lifted it to camera with the shape kept; this was the best clip of the set. Seedance 2.0 Mini had the strongest mouth motion but muddled the last words ("three coats of for lunch"). MiniMax H3 kept the block but grinned throughout and peaked at 0.1 dBFS. Pixverse V6 held a fixed wide grin for the whole clip. Veo 3.1 Fast lifted the block and looked at it naturally, then from 6 seconds looked down and let the block leave the frame.
All six welder clips had a plausible shop-noise bed under the voice. None of the twelve clips produced captions, a watermark or camera UI, which was the thing we feared most.
What broke, by model
- Grok 1.5: props morph; the brand name is guessed.
- Seedance 1.5 Pro: sleepy eyes in the first second of the founder clip.
- Seedance 2.0 Mini: scene drift; garbled last words.
- MiniMax H3: soft picture; audio at 0 dBFS, a clipping risk.
- Pixverse V6: smoothed face; a fixed grin.
- Veo 3.1 Fast: eye contact lost at the end of an 8 second clip; the most expensive per second.
Two notes that apply to every model. The mid-action in the prompt ("takes a sip") made every founder clip show a man drinking and talking at the same time, which the owner flagged on each one; the fix is a rule that nothing covers the mouth during a line, and mid-actions go in clips with no dialogue. And unknown brand names get pronounced by guessing ("ultimate", "Ultraman", "Altum"), so the pronunciation has to be given as a direction before the quoted line.
Our pick
Seedance 1.5 Pro as the default at $0.175 per clip: correct words in both sets, lips on the words, the prop handled best of the six, the highest bitrate output, 9:16 set explicitly, 44 to 50 seconds to generate. Keep the eyes open in the prompt.
Grok Imagine 1.5 for talk-only scenes at $0.11: cheapest and fastest, with loud clear audio and a face that stays itself, as long as there is no prop in the hand to melt.
Veo 3.1 Fast as a hero clip when the budget allows: the most natural acting and exact words at $0.30 for 8 seconds, which fits a 20 second reel less neatly than 5 second beats.
MiniMax H3, Seedance 2.0 Mini and Pixverse V6 are not in the pipeline. They are not bad; they lose on consistency for the money.
Since this test, every talking clip in our UGC pilots has been Seedance 1.5 Pro. Two full pilot reels built on it cost $1.75 in generations, including two retakes; the pipeline is described in how AI UGC reels work.
What this means for your reels
If you are choosing a model for talking-person reels, test it on your own start frame with a scripted line and a prop, not on the vendor's demo. Price per clip is the easy part; what separates models is whether the prop keeps its shape, whether the face keeps its identity, and whether the audio comes out at a sane level. Budget $2 to $3 for the test; it is cheaper than one wrong reel.
Then remember what the clip is for. Based on our analysis of 4.5 billion TikTok posts, the hook, the length and the caption move views far more than the rendering model. A perfect Veo clip with a weak opening sentence gets the same 300 views as a cheap one. The post on what one AI reel really costs puts the clip price next to the rest of the reel.
Ultim uses this test's winner for its UGC style and no generated video at all for its motion design style. To see what it would open your reel with, paste your site into /tools/reel-hooks.
FAQ
Which image-to-video model is cheapest with speech?
In our test on KIE, Grok Imagine 1.5 at $0.11 per 5 second clip, then Seedance 1.5 Pro at $0.175. Veo 3.1 Fast was the most expensive at $0.30, but for an 8 second clip.
Which model had the best lip sync and acting?
Veo 3.1 Fast gave the most natural performance and exact words. Seedance 1.5 Pro had correct words in both sets, lips on the words, and handled the prop best. All six moved the lips on the words.
What does a 20 second talking reel cost in clips?
At 720p with audio: Grok 1.5 $0.45, Seedance 1.5 Pro $0.70, MiniMax H3 $0.80, Seedance 2.0 Mini $0.82, Veo 3.1 Fast $0.90 for 24 seconds, Pixverse V6 $0.96.