The shortlist
| Model | Why shortlist it | Example provider price | Available in Dexto |
|---|---|---|---|
| MiniMax H3 Max | Strong published quality at a temporary low price | $0.04/s at 1080p; launch promotion | Yes |
| Wan 3.0 | Current frontier quality comparison | $0.20/s at 1080p | No |
| MiniMax H3 | A baseline against the post-trained Max variant | Check endpoint quote | Yes |
| Seedance 2.0 | An established benchmarked Seedance version | About $0.3034/s at 720p | Yes |
| Seedance 2.5 | Longer takes; evaluate the newer version separately | Token-based; see worked example | Yes, text-to-video |
| Veo 3.1 Fast | A predictable audio-on price at 1080p | $0.15/s with audio | Yes |
| Wan 2.7 | A lower-cost Wan option already in Dexto | $0.15/s at 1080p | Yes |
| Kling O3 Standard | A separate audio-on/off cost choice | $0.084/s silent; $0.112/s with audio | Yes |
| Vidu Q3 Turbo | Inexpensive low-resolution drafts | $0.035/s at 360p or 540p | Yes |
| LTX-2.3 Fast | A practical iteration baseline | $0.06/s at 1080p | Yes |
Rates are shown with their resolution and audio settings. The Dexto column lists available models; see the model catalog for current options.
What the benchmarks tell us
The Artificial Analysis text-to-video leaderboard uses blind pairwise preferences. Here is the with-audio snapshot for matching models in this article:
| Evaluated model | Elo | 95% interval |
|---|---|---|
| Wan 3.0 | 1237 | ±10 |
| MiniMax H3 Max, post-trained by fal | 1235 | ±10 |
| MiniMax H3 | 1227 | ±7 |
| Dreamina Seedance 2.0 720p | 1222 | ±5 |
| Wan 2.7 | 1106 | ±7 |
| Veo 3.1 Fast | 1086 | ±5 |
| Vidu Q3 Turbo | 1028 | ±9 |
| LTX-2.3 Fast | 978 | ±6 |
Wan 3.0 and H3 Max have overlapping confidence intervals. This table reports listener preference for the exact listed versions. Source: Artificial Analysis, September 8, 2026.
For image-to-video, use the separate image-to-video leaderboard. Starting with a reference frame changes the task. A good text-to-video result is not evidence that your product image will survive animation intact.
1. MiniMax H3 Max
H3 Max is our first candidate for a small evaluation batch because its published quality signal and current price make that experiment inexpensive. It is fal's post-trained H3 variant, so keep it separate from the original H3 when recording results.
The fal endpoint lists $0.0125 per second at 480p, $0.02 at 768p, and $0.04 at 1080p. Those are launch prices. The page says the promotion ends September 14, after which the respective rates become $0.05, $0.08, and $0.16. A ten-second 1080p clip is therefore $0.40 at the promotional rate versus $1.60 at the announced regular rate, before any platform charges.
In Dexto, H3 Max has text-to-video and image-to-video presets. Try it on one simple action, one camera movement, and one dialogue shot. Check whether the action completes and whether the sound matches the visible event.
2. Wan 3.0
Wan 3.0 scores 1237 in Artificial Analysis’s with-audio text-to-video evaluation. Its fal endpoint offers 480p, 720p, and 1080p output.
The Wan 3.0 endpoint on fal lists $0.05 per second at 480p, $0.10 at 720p, and $0.20 at 1080p. That makes a ten-second 1080p generation $2 at the listed provider rate.
3. MiniMax H3
The original H3 belongs beside Max if you want to understand what the variant changes on your prompts. fal's H3 page describes five-to-fifteen-second text-to-video generation and high-resolution output.
Dexto includes both text-to-video and image-to-video H3 presets. Keep duration, resolution, and prompt-expansion settings consistent when comparing them with Max. Otherwise you are changing several things at once.
4. Seedance 2.0
Seedance 2.0 is a useful choice when you want an established version with a published benchmark entry and both text-to-video and image-to-video access in Dexto.
fal quotes approximately $0.3034 per second at 720p and $0.682 at 1080p. Billing is based on video tokens, so those per-second figures are estimates tied to frame dimensions. A ten-second 720p example is about $3.03 using the quoted rate.
For a product animation, begin with the approved still image and describe the motion you need. Review the label, shape, and proportions across the whole clip. The point of providing a reference is to preserve the important details, not simply to get a recognizable first frame.
5. Seedance 2.5
The newer Seedance version is worth testing when your scene needs a longer continuous take. Dexto offers Seedance 2.5 for text-to-video generation.
The Seedance 2.5 page publishes token-based pricing. Its worked example gives approximately $2.31 for five seconds at 720p, 16:9: 108,000 tokens at $0.0214 per thousand. Use the token calculation for a specific request; the page's rounded per-second estimates are not an exact quote for every shape.
6. Veo 3.1 Fast
Veo 3.1 Fast is straightforward to budget for a short clip with sound. At 720p or 1080p, fal lists $0.10 per second without audio and $0.15 with it. Eight seconds with audio is $1.20 at that rate.
Dexto includes text-to-video and image-to-video presets for the Fast version. Specify whether the soundtrack should contain speech, ambience, or music. Then review it with headphones as well as watching the clip.
If the shot will sit under a separate voiceover, generating speech inside it may add work. Decide what audio belongs in the source clip before paying for several versions. Our audio model comparison covers separate narration and music.
7. Wan 2.7
Wan 2.7 is already available through Dexto's text-to-video and image-to-video presets. It is a sensible comparison when a workflow is already built around those controls and you want to judge the benefit of a newer model.
fal lists $0.10 per second at 720p and $0.15 at 1080p. Ten seconds at 1080p is $1.50 at the provider rate.
8. Kling O3 Standard
Kling O3 Standard is another Dexto option for both text-to-video and image-to-video. Its endpoint price makes the audio choice visible: $0.084 per second without audio or $0.112 with it. Five seconds with sound is $0.56.
Use it as a controlled alternative on the same shot. If you plan to replace the soundtrack, evaluate the silent version as well.
9. Vidu Q3 Turbo
Vidu Q3 Turbo is a candidate for inexpensive rough cuts. fal charges $0.035 per second at 360p or 540p, with a 2.2× multiplier at 720p or 1080p. That is $0.077 per second at the higher resolutions, or $0.77 for ten seconds.
Dexto has text-to-video and image-to-video presets. Start at the low resolution to check composition and whether your prompt expresses the intended action. Move to the delivery resolution for the final comparison.
10. LTX-2.3 Fast
LTX-2.3 Fast gives you another iteration baseline in Dexto, with both text-to-video and image-to-video presets. fal's current price is $0.06 per second at 1080p, $0.12 at 1440p, and $0.24 at 2160p.
That makes ten seconds at 1080p $0.60 at the provider rate. Compare it on simple establishing shots or a clearly described movement before assigning it a complex sequence.
Compare cost per usable shot
Take three representative prompts and generate three attempts per model. Use the same duration, resolution, aspect ratio, references, and audio setting wherever supported. Keep the prompt fixed for the first pass; record any model-specific adjustments separately.
| Check | What to look for |
|---|---|
| Instruction following | Did the requested action happen, in the right order? |
| Continuity | Did faces, hands, objects, and labels stay consistent? |
| Motion | Does movement look plausible across the entire shot? |
| Sound | Are dialogue and effects intelligible and synchronized? |
| Delivery | Is the clip the correct length and shape, with enough room to edit? |
| Workflow | How long did the accepted result take, including retries? |
Then calculate total generation spend ÷ accepted shots. As a hypothetical example, ten $0.40 attempts with two keepers cost $2 per usable shot. Ten $0.80 attempts with eight keepers cost $1 per usable shot. The more expensive generation wins on the outcome you actually need.
In Dexto, ask the agent to prepare a shot brief, generate alternatives, and save the selected clips with your project. Put recurring style and shot instructions in a skill so you can reuse them.
Which should you start with?
For a first comparison in Dexto, try H3 Max, Veo 3.1 Fast, and one cheaper drafting option such as Vidu Q3 Turbo or LTX-2.3 Fast. Add Seedance if your brief benefits from its specific version and controls. Include Wan 3.0 externally if you want a frontier reference.