9 New AI Video Models, and What They Actually Cost

Short version: nine new video models went onto the shelf this week. The two worth your attention first are Veo 3.1 Lite, because it is a Google model at roughly an eighth of what full Veo costs, and MiniMax H3, because it does 4K for about half of what anyone else charges. The rest are good, and most of you will never need them.

Quick flag before anything else: Leaxor is mine, so read the pricing bits with that in mind. The wholesale numbers below come off fal's own model pages and you can check every one of them.

What actually shipped

fal carries 136 text-to-video models. I counted this morning. Most of that number is old versions, LoRA variants and research drops, which is why "we support 136 models" would be a marketing line rather than a useful fact.

Four of them are genuinely new. Google put out Gemini Omni Flash 1.1 on 27 August. MiniMax put out H3 Max on the 26th. Alibaba shipped Wan 3.0 and Wan 3.0 Prime on the 24th. All four landed in the twelve days after I last audited the lineup, which is a decent illustration of how fast this shelf turns over.

Then there's Flux 3. Black Forest Labs make Flux, which you probably know as an image model. Early August they shipped a video model, and it runs to 20 seconds, which is longer than anything else here.

The prices

Per-clip cost on Leaxor, at each model's default settings. Credits are 10 cents each. Every upstream rate these derive from is published on fal's model catalogue.

Model	Maker	Clip	Credits	Cost
Veo 3.1 Lite	Google	4s 720p, audio	10	$1.00
PixVerse V6	PixVerse	5s 720p	12	$1.20
MiniMax H3	MiniMax	5s 768p	15	$1.50
Flux 3 Draft	Black Forest Labs	5s 720p, audio	15	$1.50
MiniMax H3 Max	MiniMax	5s 768p	20	$2.00
Wan 3.0	Alibaba	5s 720p, audio	25	$2.50
Gemini Omni Flash 1.1	Google	5s 720p	25	$2.50
Wan 3.0 Prime	Alibaba	5s 720p, audio	35	$3.50
Flux 3	Black Forest Labs	5s 720p, audio	43	$4.30

For reference, the same 4-second clip on full Veo 3.1 is 80 credits.

Veo 3.1 Lite

This is the one I'd look at first.

Google's pricing on it is $0.03 per second at 720p silent, $0.05 with audio. Full Veo 3.1 is several times that. On my shelf the comparison lands at 10 credits against 80 for a similar clip, and an eight-fold price gap between two models with the same name on the box is unusual enough to be worth a test of your own.

Durations are 4, 6 and 8 seconds. Two aspect ratios, 16:9 and 9:16. No 1:1, which matters if you were planning square.

MiniMax H3 and H3 Max

H3 is the interesting one, and not for the reason MiniMax would pick.

It does 4K at $0.16 a second. Gemini Omni Flash 1.1 charges $0.30 for the same resolution and Kling's 4K endpoint charges $0.42. Roughly half price against the nearest competitor is a real gap, and 4K matters more than people admit once you start cropping a 16:9 render down to a vertical frame.

H3 Max is the post-trained variant. Better prompt adherence, MiniMax say. It only goes to 768p.

One warning attached to H3 Max, and it's the reason I went looking at billing pages instead of taking the API at its word. fal's own note says the launch price is a promotion: "these are promotional launch rates, 50% off for a limited time. The discount ends September 1, after which 480p is $0.05/second and 768p is $0.08/second."

Today is 30 August. Anyone who wired that model in from the sticker price this week is charging half of what the render costs from Tuesday. I priced it at the post-promo rate, so nothing changes here on the 1st.

Wan 3.0 and Wan 3.0 Prime

I already had Wan 2.5 on the shelf. The thing worth knowing is that Wan 2.5 has been sitting on an endpoint literally named wan-25-preview since September 2025, and preview endpoints are what get retired when the real one ships. The real one has now shipped.

Same price at 720p. Audio on by default. Accepts up to 30 seconds.

Prime is the same model with more detail, at $0.14 a second against $0.10. I have no basis yet for telling you whether the extra 40% buys 40% more.

Flux 3

The odd one out here: Black Forest Labs shipped one model spanning image, video, audio and robot action rather than a video model alone.

Black Forest Labs' first video model, and the only one here that runs to 20 seconds.

At $0.17 a second for 720p it is the most expensive of the nine, and a 20-second clip at 1080p costs $5.80 wholesale. The draft tier is the useful part: $0.06 a second, same duration range, meant for blocking out a shot before you pay for the real render. If you have ever burned four rerolls getting a camera move right, that is the tier you want.

Gemini Omni Flash 1.1 and PixVerse V6

Gemini Omni Flash 1.1 spans 360p to 4K, from $0.03 to $0.30 a second. Three to ten seconds. 16:9 and 9:16 only.

PixVerse V6 is the cheapest thing on the shelf with sound, at $0.035 a second for 540p with audio. It also carries five baked-in styles: anime, 3D animation, claymation, comic and cyberpunk. Those are exposed in the picker.

What broke when I wired these up

This is the part I did not expect, and it is the most useful thing in this post if you are building against fal yourself.

Every video model I shipped before August declares duration as a string. You send "5", with quotes. Kling does it, Seedance does it, Veo does it, Wan 2.5 does it. So the request builder just always sent a string and nobody had to think about it.

Seven of these nine declare duration as a JSON integer.

Send "5" where the schema wants 5 and fal rejects the request. If I had trusted the pattern instead of reading each schema, seven of the nine would have failed on the first real render, in production, for a paying customer.

The audio flag has the same problem in miniature. Most models call it generate_audio. Wan 3.0 calls it audio. PixVerse calls it generate_audio_switch and defaults it off where the others default it on. Three names for one switch across nine models added on one afternoon.

So each model now carries its own answer to both questions, and I validated every generated request body against fal's published schema before any of this went near the shelf. All nine pass. That is not the same as saying all nine make good video, which is a different test and one I have not run yet.

What I haven't done

I have not render-tested these nine. The prices are verified, the request shapes are verified against fal's schemas, and the quality ranking is not mine to give you yet.

What I can tell you is what a comparison costs to run. I rendered a 5-second 16:9 clip on Kling 2.6 Pro this morning for the homepage. It cost $0.35 wholesale, took 178 seconds to come back, and arrived as a 7.6MB file. The first render was also nearly monochrome despite a prompt asking for iridescence, and I got the colour back by grading it afterwards rather than paying for a second roll.

That last part is the honest lesson from a day of this. A model that gives you a technically correct render of the wrong thing costs the same as one that nails it.

Nine new models are in the picker now. Pick two, run the same prompt through both, and keep the one that surprises you.